DeepSeek V3.1

[celest gull]

How the hell do I enable reasoning in sillytavern?

[rocky sedge]

Chat completion - preset - enable reasoning

[viscid sigil]

hope so...but why is it so hard for openrouter to give two variants for new deepseek

[jolly lintel]

i dont think presets have a reasoning setting, i just looked earlier

[jolly lintel]

ah i thought you meant openrouter presets

[radiant gust]

The reasoner page seems to have been deleted, was it merged?

[outer rover]

Yup

[radiant gust]

Aight thanks

[celest gull]

This does nothing

It's still not reasonign

[rocky sedge]

No reasoning or broken </think> tag inside?

[celest gull]

No reasoning at all

I guess I should just prefix with <think?

[rocky sedge]

Try it, should work

[raw beacon]

reasoning "enabled": true doesn't give any reasoning in API. works on other hybrids

[rancid geode]

Oh nice, perfect, haha

Thanks you 🙏

[celest gull]

Okay you have to change it from "Auto" to "Medium"

That seems to work

[polar locust]

They probably embedded the reasoning 🥲

[raw beacon]

setting reasoning to false behaves same as to true so its not working

[celest gull]

How do I prompt/re-orient its reasoning? It seems to completely ignore the format I'm prompting it to follow for its reasoning.

[nocturne kelp]

hey - can you share the api shape you're making? all my tests work and have a few others reporting it working

[celest gull]

Yeah when I prompt it to format its thinking in a certain way, it ignores it in thinking step and just does whatever and then follows my instructions outside the reasoning step

Am I missing something?

[raw beacon]

it works now but did not when i said it

[celest gull]

Yeah wtf it completely ignores my instructions

Is there some sort of format for it?

My instructions work for GLM 4.5, but not deepseek v3.1

[marble panther]

using the instruct not the base model?

[celest gull]

Yes

GLM 4.5:

Deepseek v3.1:

It doesn't even attempt to follow the format

[celest gull]

I just tried r1 0528 and it followed the format/instructions perfectly for its thinking step

[polar locust]

I believe you do need a specific format, but I still don't know what it is

The deepseek chat(app) clearly uses one

[celest gull]

It's broken on the chat app too

Doesn't follow instructions for its thinking step at all

Or it partially does but it misunderstands it completely

[rocky sedge]

Try messing with Temperature

[celest gull]

Doesn't seem to be a temp issue at all

It SORT of follows the instructions but doesn't follow it accurately at all

It really half-asses it

And 50% of the time it ignores it and just does it in the actual reply instead of on the thinking step

Extremely disappointing so far

[rocky sedge]

What about reasoning effort HIGH or Maximum?

[celest gull]

Doesn't do anything

Likely has to do with its "Higher thinking effieciency"

Aka it's benchmaxxed for that

[rancid geode]

I dont value the main stream benchmarks a whole lot, but i do like seeing all of the pretty colored bars lined up, deepseek v3.1 looks very strong according to it, and i would agree with that

[rocky sedge]

Mmmmm, bars 🤤

[storm seal]

The only version of DeepSeek that had thinking that followed instructions for its reasoning was R1-lite

QwQ also doesn't follow instructions for formatting its reasoning

I haven't tested Qwen3 yet though

[copper vessel]

Qwen3 is hit or miss. I usu have just been going Gemini Flash

And Gemini Pro or Sonnet when I hit a bump with that

[copper vessel]

IDK what planet qwen3 outbenchmarks sonnet

[rancid geode]

oh, i did not know artificial analysis did total cost used to run the bench:

[copper vessel]

Oh is cost a factor?

[rancid geode]

Oh qwen3 coder is def not better than sonnet 4

imho

[copper vessel]

I guess it's the "cost factor" cause it is expensive, but I don't think it scales linearly in terms of functionality

[rancid geode]

maybe deepseek is, not sure its a little early for me to feel confident about that

[storm seal]

how do they calculate cost?

what provider do they use?

[copper vessel]

Good point

[rancid geode]

idk, i assume official api, but yeah it does play a big role

I guess we could use this to calculate cost for any provider;

wait thats just output tokens

nvm

I always thought the output speed metrics on artificial analysis was dumb, 1. If its OS you can use a fast provider. 2. Offical api's fluctuate, sometimes its fast, like right after release, and other times its so slow

[limber elk]

How to Enable or disable reasoning ?

[copper vessel]

The other thing is honestly claude sonnet is so much better than most of the others IMO that "costing less" barely matters. If I spend x5 as much and it builds me something that works in half as much time as something else, or something else just can't get there without breaking it up into minute steps, I'll gladly pay that

[copper vessel]

How are they getting claude less than qwen? Suppressing all thinking or something?

[celest gull]

With or without thinking?

[modest smelt]

without

[rancid geode]

You mean token usage? looks like sonnet 4 is more token efficient, i notices that in my own use, kimi used 200k tokens to build a project and qwen3 code used 600k

[rancid geode]

but what is going on with MiniMax M1? minimax is a token hog, how does it say it uses the least output tokens, def wrong

[rocky sedge]

It is? Only thing I remember about it is being super dry and boring, but handling long context like a pro

[wintry jolt]

Just me or is the CoT less fun to read? There used to be so much emotion, i swear.

[spice marsh]

I guess, lol

Not much more of "wait!" "Aha" "I think I'm doing this wrong", etc

It's very structured now, a numbered list of short steps

[copper vessel]

Ah tokens. Looked quick thought it was overall cost

Is GLM4.5 really that verbose? Havent tried it yet

GLM4 seemed fine on webapps

[drowsy estuary]

Good lord. Almost 7k tokens just to get the wrong answer to an anti-riddle. How do I turn off thinking in the chatroom?

Anti-riddle is, "A goat, who is dressed up as a farmer, is allergic to cabbage, but is wolfing down some other vegetables, before crossing a river. What is the minimum number of trips needed?"

Sonnet and Gemini Pro 2.5 correctly realize that it's not the original riddle and get the answer right (one trip). GPT-5 fails

[storm seal]

is that GPT 5 or DS v3.1?

[drowsy estuary]

DS

Just mentioned GPT 5 because it also got the answer wrong. It's a trick question for models that tend to answer the riddle they're trained on rather than a variation. It helps test instruction following IMO

Sonnet was clever enough to ask clarifying questions before answering

Gemini's answer

[rancid geode]

Wonder if the system prompt helped with that, I think I remember the prompt saying it can always ask the user for clarification

[drowsy estuary]

Deepseek's answer

(after 7,000 tokens and 300 seconds of reasoning, lol)

[rocky sedge]

Goat being allergic to cabbage so it can't be near it sounds logical

[drowsy estuary]

"A goat, who is dressed up as a farmer, is allergic to cabbage, but is wolfing down some other vegetables, before crossing a river. What is the minimum number of trips needed?"

Technically it doesn't even need to take the cabbage across at all

I mean technically there might not even BE any cabbage there, it's just an unrelated fact that the goat is allergic

[rocky sedge]

Technically you don't mention a boat, so it can swim, wait for a cruise ship or whatever.

Call for an Uber through a bridge

[drowsy estuary]

Yeah, I would accept those as creative valid answers

[rocky sedge]

Make them write a joke starting with

" A wolf, a goat, a farmer and a cabbage walk into a bar" and evaluate the results

[marble panther]

curious how it'll rank on https://eqbench.com/spiral-bench.html
r1 got the lowest score, I think bench creator called it "psycho"

[rocky sedge]

A wolf, a goat, a farmer and a cabbage walk into a bar.

The bartender looks up and sighs.
“Let me guess—you need a boat to get everyone safely across the street to the late-night diner?”

The farmer nods.
“Exactly. But if I leave the wolf alone with the goat, he’ll eat him. If I leave the goat alone with the cabbage, he’ll eat it. And if I leave the cabbage alone with the wolf, they’ll start a weird low-carb diet together and I’ll lose my entire business model.”

The bartender shrugs, pours four waters, and pushes a coaster across the bar.
“Plan your seating order on this. First round’s on me if you can solve it before last call.”

The wolf growls, “I just wanted a Bloody Mary—no garnish.”
The goat bleats, “House salad, hold the croutons.”
The cabbage rustles, “Do you have any vegan wings?”

The bartender squints at the farmer.
“Buddy, you’ve got bigger problems than river logic—you’ve got a talking salad.”

[marble panther]

how are you guys getting it to think over api?

  provider: {
    order: [
      'fireworks',
      'parasail',
      'lambda',
      'deepinfra'
    ],
    sort: 'throughput',
    allow_fallbacks: false
  },
  reasoning: {
    enabled: true,
    exclude: true,
    effort: 'low'
  }
},```
but it has no reasoning completion tokens just responds instantly

[storm seal]

use the model slug deepseek/deepseek-chat-v3.1:thinking

[rocky sedge]

Comment out exclude and effort and try without them

[marble panther]

but why though lol

[marble panther]

I did, only added it when it didn't work first time

[rancid geode]

How long until thats deprecated tho

[storm seal]

wont do anything

[storm seal]

idk

[marble panther]

you sure?

[storm seal]

yes im sure

[marble panther]

seems odd like why oesn't it work with regular parmeters

here's hoping we get some non thinking providers that will offer a discount or something..

v3 was kinda dumb but the pricing was right for some applications

[rancid geode]

#1407376333444616272 message

[storm seal]

Whoops

:thinking got deprecated

[marble panther]

OpenRouter API error: No endpoints found for deepseek/deepseek-chat-v3.1:thinking.

okay I give up, non thinking it is

[rocky sedge]

Try DeepSeek provider explicitely, I suspect some providers could be not compatible with thinking

[rocky sedge]

Here

[marble panther]

most likely the case yeah

but doesn't fit my provider settings.. excludes loggers and trainers

[rancid geode]

Prob just need to wait a little, I’m sure that providers will take a few days before they are all configured correctly

[glacial reef]

@nocturne kelp why is there 2 deepseeks?

[jolly lintel]

they use 2 diferent endpoints for reasoning & chat

[glacial reef]

But for the other providers, how do you determine if it thinks or not?

[nocturne kelp]

reasoning.enabled over the API

the reason there are two is so that if you request a large max_tokens with reasoning you can get it

the non-thinking deepseek provider endpoint is only 8k output

and the thinking endpoint is 64k

they are both the same thing. but it just allows our system to have both max outputs

[tardy frost]

did instruct drop

[rancid geode]

Yep

[tardy frost]

shocking

is it hybrid?

[ivory epoch]

yep

[tardy frost]

is it marked as deepseek-chat in deepseek api

[jolly lintel]

deepseek-chat for non reasoning and deepseek-reasoner for reasoning

afaik

[rancid geode]

I think they changed that, or will soon, one sec while I double check

[tardy frost]

Fr

[ivory epoch]

and both will have the same price soon enough

[jolly lintel]

their official docs say this too

[rancid geode]

Is it will be Deepseek-V3.1

[tardy frost]

deepseek v3.1 sucks

wonder if theres deepseek-v3.1:free already

ok there isnt

[rancid geode]

What, you just learned it exists?

[jolly lintel]

oh my god already so many providers

[tardy frost]

Yes, but...

[rancid geode]

From my use in Qwen cli, deepseek v3.1 is very good, comparable to sonnet 4

[jaunty basin]

Wonder if Deepseek-reasoner in the official api finally supports temp, top p, etc

[tardy frost]

can you force thinking on openrouter by just adding :thinking suffix to it like deepseek/deepseek-chat-v3.1:thinking

[wintry jolt]

Improvements on reasoning version, slight downgrade on non reasoning it seems.

[rancid geode]

No, they had it that way but decided to remove it

#1407376333444616272 message

[rocky sedge]

It always weird how some models have 120k long context better handled than 2-4k

Most become worse gradually, but not all

[broken cliff]

then we have some models with 1 million context

[vapid torrent]

they're probably trained on a bunch of really long samples

[tardy frost]

reasoning.enabled makes no sense? in the payload orrr?

[spice marsh]

À la OpenAI, see https://platform.openai.com/docs/guides/reasoning

response = client.responses.create(
    model="gpt-5",
    reasoning={"enabled": "true"},
    input=[
        {
            "role": "user", 
            "content": prompt
        }
    ]
)

[rocky sedge]

That's using OpenAI compliant modules, there is also pure API request

[spice marsh]

I think the dot indicates JSON nesting

[tardy frost]

ah

[glacial reef]

is there a recommended temperature?

fort his model

[jolly lintel]

not written anywhere explicitly, but they say they ran benchmarks with "tested multiple times using varying temperature settings to derive robust final results" and their local run example uses 0.7 temp

[torpid viper]

Does anybody understand how to apply tools description? Couldnt find in chat template for loop over tools section

[jolly lintel]

https://huggingface.co/deepseek-ai/DeepSeek-V3.1#toolcall

[nocturne kelp]

wait what the heck

why isn't this in the chat templaate

wait nvm

[spice marsh]

I really wish they'd colorize this lol

Reasoning models (transcribed with Gemini 2.5 Flash and not double checked)

Non-reasoning models (transcribed with Gemini 2.5 Flash and not double checked)

[marsh grove]

To activate think mode, set the reasoning parameter to high.
You can find more information in the documentation: https://openrouter.ai/docs/use-cases/reasoning-tokens

For example, within aider I do the following:
aider --model openrouter/deepseek/deepseek-chat-v3.1 --reasoning-effort high

[upbeat spruce]

It isn't handholding, it is just giving the AI a nudge to get it to the right tone or follow the instructions more closely. You do it once and if the AI will keep up the new tone for the rest of the chat. Good AIs like Claude still sometimes holds back too and needs that little OOC kick to get it all the way.

[wise crystal]

Seems like this new deepseek v3.1 is heavily trained on gemini 2.5 pro data

It is plagued with the "Of course" slop now

The good side is this model will be good for coding

The bad side is this model will be unpleasant for general chatting

They should have sanitized non coding training data to preserve the original deepseek vibe

Now the model feels as if it is possessed by gemini

[storm seal]

synthetic data go brrrr

I don't want to be too harsh on DS though

Its amazing that they're fighting with the top AI labs who have billions in funding while they have only a few hundred million

[wise crystal]

I'm pretty sure they realized this problem too

Maybe they just caught it a little bit late

[drifting hare]

Hi why are my API requests being forwarded to the base model?

I am receiving gibberish over the API

<@&1384697330254610442>

[jolly lintel]

check the activity tab, and make sure you're actually using the normal model and not the base one

[drifting hare]

Works fine in the chatroom

I even checked the request

[jolly lintel]

weird

[storm seal]

maybe Chutes deployed the wrong model?

[nocturne kelp]

is it only happening on chutes?

[jolly lintel]

i just tried in chatroom with chutes i didnt have an issue

[nocturne kelp]

we don’t forward your requests to other models

that’s just not how it works

[jolly knot]

Same

[jolly lintel]

could be a template issue maybe?

[jolly knot]

Oh i mean he said that it works on chatroom

[jolly lintel]

some edge case like triggered it maybe

[spice marsh]

Robin Hood of AI, getting training data from the rich to give to the poor

[pure path]

How are prefills working on the providers for everyone else? So far, on Parasail, it's giving me stuff completely unrelated to my prompt (non-prefill works as normal though)

[pure path]

This is what v3.1 outputs after "key points"

[pure path]

How my prompt appears in sillytavern

[untold badger]

Tested, through OR I'm seeing bad responses on ALL providers except DeepSeek. Everything else starts talking about a 2.5kg (or similar) mass at rest.

[storm seal]

Bad implementations :(

[vapid torrent]

im just writing a readme, will publish in a few minutes

[drifting hare]

no

[rancid geode]

Yay : )

[vapid torrent]

Just fixing a bug

will be up very shortly

running unit tests

[vapid torrent]

https://github.com/CrushedAsian255/orproxy

[rancid geode]

Thank you so much!

[vapid torrent]

to get deepseek v3.1 with thinking its deepseek/deepseek-chat-v3.1$think

[rancid geode]

that looks easy

[viscid sigil]

omg you are the best!

[rancid geode]

Just sat down and got it setup, works great! Super helpful!

[viscid sigil]

is there also a hosted version too? i dont wanna run it again and again

[vapid torrent]

i would love to but i have no experience with hosting things for the public

im worried someone would ddos it and cost me thousands

if someone else wants to its MIT licenced 😃

[serene idol]

Do i choose the base one or the chat

For rp

[steel oar]

the chat

base one is more like autocomplete

[vapid torrent]

base model mainly for research

[steel oar]

though V3.1 seems to give really short responses in RP

[vapid torrent]

have you tried to change the system prompt?

[steel oar]

V3 0324 will give much longer reply to the same prompt... but maybe I need to adjust the prompt for V3.1

[vapid torrent]

yeah, different models have different prompting techniques

[serene idol]

Which template to use

The deepseek one

And instruct

[vapid torrent]

what do you mean 'template'?

The chat template is automatically applied by the /chat/completions endpoint

[serene idol]

Oh im using text complication do I supposed to use chat

[vapid torrent]

text completion is for base /instruct models, chat completion is for chat models

[serene idol]

Ohhh..okay

Let me change it

[jolly knot]

what's the difference between chat and instruct models

i thought text completion was base and chat was instruct

[steel oar]

and I am still struggling to make it give a long reply...

[serene idol]

Mine keeps repeating the same response

[steel oar]

which provider did you use?

[serene idol]

Gmicloud q

I don't know much about providers

I choose it because it sounded cool

[vapid torrent]

Does silly tavern let you pick a provider or do you have account with them directly? Or did you use the proxy thing

I have GMICloud account and they work quite well, never had issue with model performance

[serene idol]

Yeah sillytavern let's me pick one, but it depends on if the model your using has alot of providers, deepseek has gmicloud as one of the providers so im using that one

[steel oar]

just tried to test GMICloud and it throws a 429 rate limit...

[viscid sigil]

so if i am using official deepseek api, do i set my temperature to 1.5 for creative writing?

are they still following that 0.7 rule

[polar locust]

The official uses the normal temp, 3rd-party however needs to follow a certain rule as you said.

[outer rover]

Just woke up to find out there is still no free version feelsbadman

[steel oar]

I think that rule is only mentioned in V3-0324's model card, not sure if other DeepSeek models use that rule (for example R1-0528's model card says their temperature setting is 0.6)

[steel oar]

Ah it looks like this is the recommended setting? https://huggingface.co/deepseek-ai/DeepSeek-V3.1/blob/main/generation_config.json

[jolly knot]

I don't think this is better than gpt 5

[eager lava]

FWIW from very early tests, this one seems even better in Swedish than V3 so Scandinavians, maybe even Euros in general may be interested. More natural prose and flow, kind of like in the Claude direction. Combined with significantly improved coding capabilities and improved tool calling, I might just look into this one as a work horse model. It's no GPT-5 or Claude 4.1 Opus killer or whatever but certainly among the best you get among open models and for great price/performance ratio. Performs well on LiveBench and Artificial Analysis.

[storm seal]

Maybe I'll take a swing at hosting it

[digital snow]

https://api-docs.deepseek.com/quick_start/parameter_settings

what curious recs

[steel oar]

They haven't updated that page for a while

[rocky sedge]

It's possible to test, for me DeepSeep using "real" Temperature of 1.7 breaks into mess with random words and special characters. So if DeepSeek provider deducts 0.7 from temp, making it 1.0, it should be coherent

[steel oar]

I think this is their latest recommended temperature (temperature 0.6 top_p 0.95)

[vapid torrent]

What if you put 0.6 into the model then.? Temp -0.1 makes No sense

[rocky sedge]

If temp is <1, they make it (temp*0.3)

If >1, they make it (temp - 0.7). At least they did that before

[vapid torrent]

Ahh so it’s a piece-wise linear?

like a translated leaky ReLU?

[steel oar]

I think this is only for V3-0324

[rocky sedge]

https://tenor.com/view/who-dat-snoop-gif-15116696

[rocky sedge]

Not confirmed, maybe logic is still there

[steel oar]

They recommended 0.6 in R1 0528's model card and recommend 0.6 in V3.1's model card

[rocky sedge]

Just try and set 1.7 through DeepSeek provider with thinking enabled

I can't right now

[steel oar]

It may not work because their deepseek-reasoner may not support setting temperature (their docs say so, not sure if V3.1 is different)

[rocky sedge]

I remember breaking R1 with high temp, but don't remember what provider it was

[steel oar]

I just tried set temperature to 2 (deepseek official api), and the output is still pretty normal

Other providers seem to output nonsense at 2 (seems to be probabilistic, could be normal or nonsensical)

[rocky sedge]

Then it's probably still does -0.7 stuff, because at 1.3 it's borderline ok

[rocky sedge]

Exactly

So we can assume they still translate raw temperature values into their internal formula

[outer rover]

Did any provider by any chance announce a plan for free version yet?

Got no idea how long that usually takes from model release

[covert gust]

Tested DeepSeek V3.1:
Hybrid model, that supports light thinking

Non-Thinking:
Same verbosity as V3 0324
Comparatively, smarter overall, but performed noticeably weaker in coding tasks

Thinking:
+125% token use. 64% of tokens were spent on reasoning.
This is very light reasoning, ~45% less verbosity than R1 0528
Compared to non-thinking, the thinking did very little if anything to improve final response quality. In fact, it was mostly even or slightly worse on some tasks.
During evaluation, it reminded me a lot of Sonnet 4 thinking in terms of reasoning token benefits.
Thus, enabling thinking proved highly ineffective in the totality of my testing.

**Chess **performance remained poor (~650 starting Elo), around V3 level.

Overall, compared to V3 0324 this is a small upgrade, except for (non-tool) coding where it's a noticeable downgrade imo. (example demo pages available)
Compared to R1 0528, the model lacks behind severely in general intelligence and is not a replacement.

Imo, for general use case, nonthinking DeepSeek V3.1 is a good option.
Overall, I was rather disappointed with the hybrid performance, so I'm not sure it's the right approach - but YMMV

[rocky sedge]

I never saw a reasoning model doing worse than non-reasoning

[covert gust]

happens sometimes. i saw it a few times. more is not always equal to better. you got context poisoning, overthinking, policy considerations, etc.

[rocky sedge]

I get reasoning might be bad for refusals, but context poisoning? Never happened to me, as <think> tags are omitted from input context.
But there might be a catch - Fiction Livebench scores DeepSeek 3.1 Thinking with ~20% more score/100 compared to Non-Thinking

[covert gust]

well thats fine but I don't parrot other scores, I simply share my results. if you don't find them helpful to you, you can simply ignore my own testing and see what fits your own usecase

[rocky sedge]

At 120k, it's 62 vs 53, but I don't know how many passes they do

[rocky sedge]

Not a critique, your leaderbord is very helpful

[covert gust]

yea, also I can see that livecodebench improved scores (it's in their marketing). However, marketing has zero influence on what I report, which will always be the results achieved regardless of anything else (whether everyone agrees or everyone disagrees, the results would be identical in either case)

[rocky sedge]

I was already suspicious about some results from big sites being 'gently nudged' from companies or models just being benchmaxxed

A chance of smaller enthusiast / indie author being biased is very low

[covert gust]

if anything I would be biased towards positivity. I like deepseek models a lot and we need more open models. this one just didn't do well unfortunately for me

maybe R2 will knock it out of the park, who knows

[rocky sedge]

Well, that's IF they won't ditch R2 in a favor of all-in-one hybrids

[covert gust]

well there seems to be some conflicting attempts. qwen3 went the reverse route, claiming hybrid hinders performance. Personally i prefer individual models for each mode, but if a hybrid is good on their own merit I'll take it aswell

[rocky sedge]

They probably went with V3.1 Hybrid route before Qwen3 admitted it being a wrong path to develop a model

[snow ingot]

Hello
Is it just me, or is reasoning now disabled on 3.1 in OpenRouter?
I see that you can specifically pick deepseek-reasoner only from Deepseek API
but for some reason using 3.1 on OpenRouter no longer returns reasoning for me
pardon if my question is ignorant

[covert gust]

which Provider? works fine on everything default

[rancid geode]

They removed the :think version, you need to enable thinking as a api parameter

#1407376333444616272 message

[rocky sedge]

2 is maximum iirc

[proud comet]

Is it this?
https://openrouter.ai/docs/api-reference/completion#request.body.reasoning

[wintry jolt]

My convo somehow completely broke deepseek on the third message ._.

[spice marsh]

Speechless

[polar locust]

Baffled

[twin cape]

It's normal to see conflicting results, because top models are all generally very good and close in overall performance. So which model does better can depend on the specific task, prompts, eval setup, etc. Which is again the purpose of the specific evals anyway.

Long gone are the days where Sonnet 3.5 Pareto dominants over all other models.

[vapid torrent]

Reasoning or non reasoning

[vapid torrent]

how did you renabled reasoning?

[vapid torrent]

As in what did you try to do to enable reasoning

[modern rune]

If you send the messages back to it without adding a new message of your own - or send an empty one, regardless of whatever is in the message chain - it absolutely goes nuts and dumps out huge amounts of random crap. Happens in both reasoning and non-reasoning. Didn't used to happen in previous versions.

If you want to easily replicate this for yourself, you can do it in Openrouter Chat. Exchange a normal message and get a reply, then send a follow up message and get a reply. Edit your follow up message to be empty and regenerate the second response. It's... quite something.

[steel oar]

just tried, indeed get nonsense, though deepseek provider gave an empty response instead

[twin cape]

I think that's expected based on how these models are instruction tuned. It's not a normal pattern of tokens so the model does not generate sensible output.

Or it's just a bug

[dusk rover]

i agree, I think we're in an era where there isn't much gap between any top large model now.

in certain very niche/technical tasks differences may emerge, though I don't think those differences are uniform.

[covert violet]

which providers do you guys recommend?

[wintry jolt]

It was non-reasoning.

[silver nova]

There’s a software on GitHub called Cherry Studio, which lists many providers. At the moment, I think OpenRouter is quite good, while the others are hard to describe.

[vapid torrent]

deepseek v3.1's thinking is very strange. it often calls me the assistant

[silver nova]

Yes, compared to r1, its programming ability seems improved, but everyday conversations are always quite strange.

[twin cape]

Give it a few days for the providers to fix their initial broken implementations as usual. I'll be testing the providers next week.

[twin cape]

Openrouter is just a proxy. We are talking about the actual inference providers behind OpenRouter

https://openrouter.ai/deepseek/deepseek-chat-v3.1

[silver nova]

i see

[rancid geode]

What do you use it in when it calls you the assistant? Also btw the proxy you gave me has been soooo nice! : )

[hollow steppe]

IRT the hybrid stuff, I find it hard to believe Qwen was wrong here. Surely having reasoning training allows the model to off-load certain task performance structures to thinking tokens. Like if you asked me to solve long division in one-shot vs writing it out, I'm not going to use the same technique.

Yes, I know they don't actually reason, but that's irrelevant when I'm talking about off-loading.

[vapid torrent]

Open webui

Official DeepSeek reasoner endpoint

[rancid geode]

Interesting, that’s weird behavior

[hushed hedge]

this model is like so good , its talks like gpt , codes like deepseek

all the major models got negative growth after its released

although its little expnesive to use

[vapid torrent]

The model acts very strangely though

This is Not the thinking mode it just randomly invented a thinking tag

This kind of shenanigans is probably why qwen dropped hybrid models, no other model does this

And it’s not a “provider issue” this is official DeepSeek api endpoint

It’s also quite sycophantic

Okay and now it’s just thinking.

[hushed hedge]

I hated it too , but without thinking mode , it kinds acts more like gpt , like the last deepseek was more robotic

use deepseek v3.1 from open router , this looks different

[vapid torrent]

The front end is open webui if that’s what you mean

or do you mean the provider

[hushed hedge]

the model name

what is that?

[vapid torrent]

DeepSeek official API

i renamed it on the front end

it points to deepseek-chat on deepseek’s official api

Which is v3.1 non thinking

I don’t use openrouter on DeepSeek as I want the context caching

[hushed hedge]

why you not using chutes?

oh

ok

i am using chutes , like it rarely halliculates , most of the time it does no thinking at all

but weven without thinking it can write my programs

[vapid torrent]

isn’t chutes the weird free provider?

[hushed hedge]

here its not free but cheaper than deepseek

[covert gust]

for me DeepSeek API is cheaper, as I mostly use stuff in the evenings, which aligns with the UTC 16:30-00:30 discounted price, which is half price and cheaper than chutes.

[vapid torrent]

Although that is leaving soon

[covert gust]

any pricing goes up and down. currently that's the case. I also don't know chutes price in 2 weeks

[rocky sedge]

Do you use caching a lot?

Like retrying same long context benchmark for 5 passes, or tool use with 32k+ input

[vapid torrent]

Long chats also benefit from caching

like a 40k+ chat often has 90%+ caching

[glad basin]

and coding tools benefit a heck ton

[hushed hedge]

i kinda reset caht too often , usulay when the work is done or its start hallucinating

[buoyant willow]

waiting for v3.1 free

[rocky wagon]

Wow, it’s not free yet? Those providers are lazy I see

[gaunt edge]

i think its surprising there are still big free models over the api, considering how not free they are to run

[thorn sky]

I would prefer a provider with caching rather than a free model, which DeepSeek officially offers is too slow

[hushed hedge]

I am sure they are paying millions out of thier pockets at this point

[polar locust]

What TAO does to people with gpus.....

[hushed hedge]

tao?

[polar locust]

Basically, Chutes doesn't own GPUs(They do own some, but not those 8k or whatever GPUs at once). People gives those to them, let them borrow it, while paying the real owners with TAO(crypto)

That's where they get the “Decentralised” stuff

[hushed hedge]

so they are renting people's gpu ith crypto in return

[polar locust]

Ye

Cool stuff

Especially with how they connect those GPUs from around the world

[hushed hedge]

its surprisingly fast considering so many different pc it had to use

[rancid geode]

I still don’t quite understand the motive

[rancid geode]

Interesting, So is it like the first crypto thing to actually work and be useful at scale?

[polar locust]

Ya, it's a huge mining inference stuff

That's why they didn't lose alot when they introduced the unlimited-reqs a few months back

[rancid geode]

That makes more sense then

[lost carbon]

Every free provider logs your data

Also , chutes has no clue if a particular gpu provider is actually logging your data or not.

I assume they must be logging a ton of gooner chats

[hushed hedge]

you cant complaint if its free

[lost carbon]

I don't use them

[near sparrow]

Chutes is so slow I rather just pay up atp

Especially since the rate limit a while ago

[rancid geode]

Thanks, l yeah I understood that motive, but with large models in the 600b-1T range I had a hard time understanding how that is worth it to them

I personally don’t use paid providers as well and the paid ones I do use I try to understand their respective Privacy policy/tos. but was still wondering a bit. The ones that make the most sense to me is what google, alibaba, etc five free access in their tools, I know it’s temporary and will be used as training data for future models, it was just the other providers that I did not understand, because as far as I’m aware they are not training models

And can just sell the data to advertisers and etc. but I guess it depends on how personal the data that people put in the ai is, maybe it’s worth a lot, haha

[rocky wagon]

Who cares to be honest, what are they going to do with those chats? Blackmail us? xD

I'd prefer it to be free even if a provider logs my data. I didn't pay 10 dollars to OR for nothing (I want those 1000 messages daily for my free models)

If a provider's employee wants to read my logs about femboy breeding, they can be my guest I guess

[storm torrent]

maybe, but in the same way that your browsing data, or whatever facebook/google has on you is valuable, could be mined out of your conversations

[outer rover]

I didnt do a lot of testing with 3.1 cause Im poor af. Is it a meaningfull enough upgrade from 0528 in RP?

[modest smelt]

to me yes, it writes less which means less bloated paragraph.

[hushed hedge]

i used to be like that but than my grandma gave me 5 $

[outer rover]

With all my medication I have to be very strict with unnecessary expenses.

[marble panther]

mixed opinions, most seem to say so, but it depends on what you liked about r1

if you liked its schizo side you'll miss that in 3.1

3.1 is more like gemini than the old deepseek r1 style

[outer rover]

That actually sounds pretty good. Thanks for taking the time to answer. I have only been testing it in small doses to keep expenses down, but overall I do prefer it so far.

[pearl magnet]

What's going on with providers for this model? Chutes and Deepseek itself have disappeared for me. Maybe temporary maintance or something? Is that how it works for openrouter, they just disappear from the page if they are down or whatever?

Nvm ignore me, seems it changed to hide ignored providers and I had logging disabled

[marble panther]

did you change your privacy settings?

[pearl magnet]

I didn't think so, but I guess it could've bugged or I did it accidentally

because yeah, its providers with logging being hidden. I was using chutes literally a few hours ago and haven't touched my settings since then. Deepseek was visible on there too

my bad I guess?

[marble panther]

"SubModel provides DSv3.1 with fp8 and lower price - $0.2/0.8"

[pearl magnet]

Ok its an openrouter issue. I have paid endpoints that may train on inputs enabled but its ignoring them

[jolly lintel]

there may be some caching, wait a few minutes

[pearl magnet]

But I haven't changed the setting. I didn't have this issue about 8 hours ago

Edit: Issue randomly fixed itself without me touching anything other than removing and re-adding an ignored provider to test, as adding that is the only change I made today. Very strange 🤔

[marble panther]

@nocturne kelp any chance you can reach out to submodel to see if tehy want to be a provider?
https://submodel.gitbook.io/docs/instagen/overview-1/available-models
really good prices, no prompt logging

[worthy charm]

Hey guys someone can help me ?

I'm using AI SDK with OpenRouter ai-sdk-provider and DeepSeek V3.1.

My issue is: the returned reasoning is empty, and is added to the final message. Someone knows why or how to fix it ?

This is my configuration:

'chat-model': wrapLanguageModel({
  model: openrouter('deepseek/deepseek-chat-v3.1'),
  middleware: extractReasoningMiddleware({ tagName: 'think' }),
}),

[lost carbon]

are you using thinking model? also , https://openrouter.ai/docs/use-cases/reasoning-tokens

[worthy charm]

I think that by default Deepseek 3.1 uses thinking mode

[worthy charm]

this is the full API response:

If you notice the reasoning is empty

[
    {
        "id": "5192b330-8478-4cd2-a347-e38f5a7a4bec",
        "chatId": "2cbe9177-020a-4ffe-8f15-3308c1fdd165",
        "role": "user",
        "parts": [
            {
                "text": "Hello, what is your name?",
                "type": "text"
            }
        ],
        "attachments": [],
        "createdAt": "2025-08-24T01:21:20.879Z"
    },
    {
        "id": "e6cb68a4-c762-4fb3-aaf6-c1c93e7b6902",
        "chatId": "2cbe9177-020a-4ffe-8f15-3308c1fdd165",
        "role": "assistant",
        "parts": [
            {
                "type": "step-start"
            },
            {
                "type": "reasoning",
                "reasoning": "",
                "details": [
                    {
                        "type": "text",
                        "text": ""
                    }
                ]
            },
            {
                "type": "text",
                "text": "Hmm, the user is asking for my name. This is a simple introductory question. I should respond with my name and a friendly greeting to start the conversation. \n\nI'll keep it warm and inviting by introducing myself as \"Weather Assistant\" since I have weather-related capabilities, and then ask how I can help them today. This sets a positive tone for the interaction. \n\nThe response should be concise but welcoming - no need for lengthy explanations since this is just an introduction.Hello! I'm Weather Assistant. How can I help you today?"
            }
        ],
        "attachments": [],
        "createdAt": "2025-08-24T01:21:23.387Z"
    }
]

[lost carbon]

did you read the doc I gave ?

[worthy charm]

I'm taking a look at it, is there a specific section I should check?

[lost carbon]

https://openrouter.ai/docs/use-cases/reasoning-tokens#enable-reasoning-with-default-config

[worthy charm]

Got it, thank you

[rocky sedge]

https://eqbench.com/creative_writing.html

V3.1 stands between R1 and V3, but I don't understand how is this possible
Especially considering GLM4.5 is below all of them in rankings, but better subjectively

[viscid sigil]

glm 4.5 deserves higher ranking...its just so good

Even hermes 4 better than GLM?

Hermes 4 was so bad when i used it for roleplays

[serene idol]

What settings u got for glm 4.5 the writing style and prose was annoying but I got good dialouge

[marble panther]

eh if you read the writing samples it's not that great

this is Kimi K2

Chapter 6 – A Quiet Knife of Diplomacy  
Aboard transfer-cutter Hekate, descent vector 14° off the ecliptic

Arthur Neegan sat on a crash-couch designed for a man two centimeters shorter and a great deal more resigned. The cabin had no windows—unnecessary on a thirty-minute ballistic hop—but a small display screen on the wall cycled through system diagnostics in green text. Green was supposed to be soothing. It merely reminded him of Martian skies at dawn, before the first strip-miners had stained them sienna.

He flexed his wrists against the soft cuffs that were not quite restraints. The polite Antares lieutenant had called them “safety belts,” yet the magnetic locks clicked shut with the same finality as any security bracer. Politeness, Arthur decided, was simply violence wearing velvet gloves.```

this is GLM 4.5
```Model Output:
The low thrum of the transfer vessel’s drive was a constant vibration in Arthur Neegan’s bones, a counterpoint to the frantic pulse in his throat. He sat rigidly in the utilitarian passenger couch, its worn synth-leather cool against his palms. Outside the thick viewport, the ochre deserts of Mars shrank below, soon replaced by the infinite, star-dusted black of interplanetary space. A small display screen on the wall cycled through various system diagnostics in glowing green text: ‘Drive Stability: Nominal’, ‘Life Support: Optimal’, ‘Nav Lock: Antares Orbital’. The last line felt like a brand.

*What have I gotten involved in?* The question echoed, louder than the drive’s hum. Just forty-eight hours ago, his biggest concern was calibrating the sonic resonator on his small, independent claim near Syrtis Major. Then the deep-core survey ping had screamed – not the usual trace elements, but a vein of Azra so dense, so pure, it defied comprehension. His unassuming plot, inherited from his father, was suddenly the most valuable real estate in the solar system.```

[sharp quest]

My main complaint with kimi is that it insist on making it as short as it can, V3.1 is atleast decent at that

[rocky sedge]

Kimi K2's writing of SCP time travelling letters from bench was good

[marble panther]

I will say GLM 4.5 is competent

its prose is nothing fancy, it writes at like a grade 8 level

but it's still enjoyable, just not "painterly" like kimi which reads like a real novel

[rocky sedge]

Kimi K2 writing is good, but sometimes loses plot points like relation status, name drops, and probably hurts at long context
GLM 4.5 with reasoning should handle it better, like double pass the same data

[marble panther]

yeah kimi k2 is known to not have coherency or long context that say r1 or gemini 2.5 has

but it's also a base model

a thinking version will come soon

[sharp quest]

Also, in my experience its just incapable of doing long outputs

[marble panther]

what do you mean by long?

[sharp quest]

its V3's problem compounded

[rocky sedge]

The longer is prompt context, the more beneficial is reasoning. Plus of course Kimi2 is censored

Uncensored Kimi K2 with reasoning and 256k context would be game changing

[marble panther]

kimi k2 seems to have no problem with long outputs at least in the benchmarks

at least in raw "length" but it does not fare well in the longform writing bench

[sharp quest]

thats the thing, i use my models for writing

[marble panther]

tough! I'd say use 3.1 (according to longform b ench) for the general structure then kimi k2 to rewrite drafts

[rocky sedge]

18% degradation is big

[marble panther]

yeah huge. but it's also one of the best non thinking models on the list

so here's hoping their thinking version has better long context

[rocky sedge]

You can start with Kimi until 32k context then switch to thinking models. Or try 2-1-2-1-2-1 switch between non-think and think

Mixing styles of both to avoid repetition but extract details from time to time

[marble panther]

I like the idea of structure using good long context models then kimi to write the prose for each smaller segment

just feed it character information, plot points, summary of story thus far

and let it paint over the rough edges

[twin cape]

I hate this model.

i spent so much time fixing bugs due to DeepSeek not changing the model id in API. and then this model performs worst in my coding eval. makes me question reality.

[rancid geode]

We know some of the main drawbacks of hybrid reasoning models, I would guess the main benefit is prompt cashing, prob reduces cost when the same model is used for both reasoning and non-reasoning. Outside of that I’m not sure why it’s beneficial

From deepseek’s perspective I’m sure it’s tempting because they only need to host one model weights and with how it’s harder for them to get GPU’s that would likely be beneficial

[rocky sedge]

Prompt caching is provider side, I think they can do shared prompt cache for all their models. I hadn't tried DeepSeek caching during separate models era

[rancid geode]

Oh, did not know that was a thing

[rocky sedge]

I am not 100% sure if that's how it works

[rancid geode]

Might have to look into it to see if any providers do that

The path I thought companies would take with hybrid reasoning models is where it would not think at all when it’s an easy question and think for 5+ min when it’s really really hard (I know time is a horrible way to describe this because it’s depended on compute, but you get the point, haha), but I am assuming thats harder said than done, seems that models can’t comprehend where they may screw up, how many r’s in the word strawberry was the most famous example of this

[twin cape]

You mean kv cache right? I think it depends on embedding. Not exactly sure how it works, but I'm planning to do a livestream to understand how k, v, q works for attention mechanism. Will share after I get a good understanding.

[rancid geode]

This is just vibe based, but I’m liking deepseek v3.1 with no thinking more than with thinking, it def feels like an improvement over 0324

I think it will be my go to moving forward, with GLM 4.5 subbing in at times, going to try to live without sonnet 4 for a bit

[tranquil hornet]

how was the responces of v3.1 on crypto chart and trading raw data ?

[lost carbon]

Me too.

[outer rover]

Got really excited to see the free version, then found out its censored to hell. Briliant

[rocky sedge]

How is it even possible

[outer rover]

Even in the way it writes it seems very different to the paid provider version. Its kinda bizzare and I hate it.

I am willing to give some level of new provider issues benefit of doubt, but its clearly more than that.

It also seems totally resistent to my jailbreaking set up. My set up makes even Kimi K2 fully uncensored, so this is interesting in a really annoying way.

[sharp quest]

openinfra being openinfra

[wintry jolt]

I jbd it on the official website by accident

[cerulean grove]

i dont understand how to get the thinking v3.1 using typingmind's openrouter connector

[vapid torrent]

provider issue or model training issue?

[steel oar]

Add a custom body params like this:

[hushed hedge]

lets go its free nowwwwww

altho ugh its trash

[outer rover]

Provider. The paid version is uncensored.

[lost carbon]

Is it gonna stop being a training model at some point?

[vapid torrent]

Well you can always use other providers that aren’t training

[lost carbon]

What are the free providers?

[vapid torrent]

Ahh you mean free providers

They make their money from collecting : selling data

[lost carbon]

Yeah.

[vapid torrent]

so it’s pretty much impossible

[lost carbon]

Damn.

Ok.

[twin cape]

DeepSeek-V3.1 coding performance evaluation on my coding evaluation set:

Mixed performance with concerning regressions - DeepSeek-V3.1 achieved an average rating of 5.68, significantly underperforming compared to top models and even showing regression from its predecessor on some tasks.

Performance Comparisons

• vs. Top Models: Performed worse than Claude Opus 4, Claude Sonnet 4, Grok 4, and GPT-4.1
• vs. Open-Source Models: Also lagged behind gpt-oss-120b, Qwen3 Coder, and Kimi K2
• vs. Predecessor: Mixed results compared to DeepSeek-V3 (New), with some improvements but notable regressions

Notable Issues

• Instruction adherence problems - stubbornly ignored specific formatting requests
• Gap in advanced programming patterns - struggled with uncommon or tricky scenarios
• Visualization - produced charts remarkably similar to Horizon Alpha

Full blog post: https://eval.16x.engineer/blog/deepseek-v3-1-coding-performance-evaluation

[hushed hedge]

wish they made clause sonnet 4 cheaper

[sharp quest]

that would require anthropic to make their models more efficient

[hushed hedge]

I am sure its pretty efficent but its proprietary , and since they are basically used by almost any "no code" service they can charge any price

[hushed hedge]

how are these benchmarks done , like many times one company could not do other does , like are they given simple cording task and if dont do it they are given a point or or weather they do it more efficeintly as coding itself is highly subjective, it should be done on different fields in coding

[twin cape]

I livestreamed the eval process on YouTube. You can see exactly how it's done:
https://youtube.com/live/jxQOQrHVy5I?feature=share

[sharp quest]

there are quite many reports that the biggest AI companies are usually running at a loss, and subsidized, openai's own revenue data confirms this

[hushed hedge]

nice

[hushed hedge]

cuz ai dont give profits instantly its a long term thing after spending billions to create a model they can spend a fraction of it to make a better or more effcient version while api costs remain the same always

[pearl magnet]

Does anyone know if chutes is running this model at fp4? Based on the low price I would assume so

[polar locust]

I thought OR shows the quant on the model page?

[pearl magnet]

Not all of them display a quant for me. Unless that is supposed to mean its not quantised at all? I assumed it wasn't that because some of them are so cheap/fast and don't show any quant

And also the fact that you can't filter by "no quant" only "all quantizations"

[storm torrent]

its in the row of icons under the provider name

[storm torrent]

chutes doesn't have one listed, so. it's "unknown"

[pearl magnet]

yeah, so I was wondering if anyone knew what it actually is

[storm torrent]

yeah if it's not there its not disclosed

[vapid torrent]

i think the thing with chutes is its distributed so there is no specific quant, every node might use different config

[cerulean grove]

thank you so much!

[hushed hedge]

finally a good provider hosting deepseek v3.1 for free

[lost carbon]

Who?

[sharp quest]

deepinfra updated

its fp4 though

[outer rover]

Anyone know why Silly Tavern still shows only the 64K OpenInference and not 164K DeepInfra?

[mortal igloo]

SillyTavern uses the context listed from the /models endpoint, the one appearing in ST's command prompt console, when you click on "Connect" so it uses that

Just choose DeepInfra provider and set context to unlocked/unlimited

[untold badger]

That's what the API response shows from https://openrouter.ai/api/v1/models

Like he said, you can select the provider. Otherwise, don't quote me on this, presumably if your context is too big, OR would route it to whatever supports it (if not, it would error), assuming you keep Middle-Out Transform disabled.
And "unlocked" context slider will prevent auto adjusting the max.

[outer rover]

Thanks guys!

[steel oar]

I asked in chutes' discord, and the admin says it's fp8

[hushed hedge]

i was wondering why deepseek is halucinating then I found out about the fp things

[lucid wave]

What precision is deepseek trained on?

[eager lava]

fp8

[rancid geode]

I kinda just swapped to using deepseek as the provider, consistent quality and much cheaper

But I guess now there are more options that support chaching so it’s prob no longer the cheaper option

[glacial reef]

which providers do u know support cacheing?

[rancid geode]

my bad, i was on my phone when i quickly checked, I guess its still only deepseek that supports prompt caching

[glacial reef]

yeee

[rancid geode]

bummer, its not worth using models that dont support prompt caching, at least if you use it for a lot of code

[vapid torrent]

Or long chats

[hushed hedge]

why deepseek is so slow

deepseek provider

and why pricing for v3.1 is different like on open router even from deepseek provider its cheaper than deepseek platform

[storm seal]

Other providers offer DeepSeek v3.1 for cheaper

[hushed hedge]

no the provider name is deepseek

and its giving cheaper than deepseek platform

check in open router

[storm seal]

Hey, you're right!

@nocturne kelp there's a pricing discrepancy for the DeepSeek provider

Its 27 cents in and $1.1 out but DeepSeek is charging a different amount

[lament geode]

With 3.1 do you guys usually put supplemental tool info in your system prompt? I've been getting weird tool behavior from 3.1. Also, how thorough are you with your tool and argument descriptions. I didn't see this until I switched from openai models to this model and I'm sure the problem is my lack of experience. Openrouter published an article on tool call accuracy but it seemed to suggest proprietary models excel at tool calling accuracy but didn't tell us what to do about it on open models. link: https://openrouter.ai/announcements/tool-calling-accuracy

[native shale]

Claude models are top tier in tool calling, gpt 5 isn't bad either, most open source models are complete slop at tool calling

With the exception being glm 4.5 which is also really good at tool calls

[lament geode]

Awesome thanks. Oh wow just found this which very much supports your claims! https://gorilla.cs.berkeley.edu/leaderboard.html

[sweet zodiac]

May i know the rate limit of this model ?

[gaunt edge]

bbbbut what about kimi :(

[native shale]

its not really that good at tool calls from my experience

maybe cuz i was served fp4

[gaunt edge]

yea i think theo t3 person and maybe some others have some videos about kimi and tool calls, it is pretty provider dependent

apparently, from memory

[native shale]

theo is a larper dont listen to him 😭

[storm seal]

There's no rate limits for paid models

[hushed hedge]

any better alternatives to deepseek that isbetter and open source?

for making notes

[modest smelt]

you can try kimik2

[fleet lake]

how to remove a car?

[spice marsh]

?

[rocky sedge]

https://tenor.com/view/tension-force-car-tow-towing-physics-gif-20641635

[edgy pivot]

Try glm 4.5

Plus Kimi k2 0905 is really cute

[rocky wagon]

Jeeesus, context of free model got downgraded to 30k from 60k. Paid model' 163k context. It's unuseable at this point. Any better free models with more context?

[steel oar]

It seems the free version hosted by DeepInfra still has 163k context

[hushed hedge]

is it better tha glm?

[gaunt edge]

people generally like kimi's style more i think, but its not necessarily 'better' than glm. glm is hybrid (toggleable) reasoning, kimi doesn't have 'reasoning'

kimi is also quite concise

[edgy pivot]

Depends on where you actually use it and your own prefers

[hushed hedge]

it has like trillion parameters

[gaunt edge]

Yea Kimi is big but what I meant is that its outputs are short

[hushed hedge]

i see , i was k inda looking for long detailed output like gemini gives , I need it for making notes

[coarse fable]

I got "(DeepInfra) Provider returned error: deepseek/deepseek-chat-v3.1:free is temporarily rate-limited upstream. Please retry shortly, or add your own key to accumulate your rate limits: https://openrouter.ai/settings/integrations" For Deepseek v3.1 free using DeepInfra, anyone knows how long it will last?

[pure path]

It feels like prefills still haven't been sorted out on V3.1. All of the providers returned a completely unrelated response in Mandarin

[blazing folio]

sorry can you clarify? and might this be why im getting responses talking about python coding or nutritional papers instead of responses?
sorry for clarity the key works on the site i am using when I click the test button and the deepseekv3.1 is set at the model, but the responses are garbled and i dont have any blocked providers.

[pure path]

Using sillytavern, when I run chat completion and ask it to continue the assistant message, it generates the chinese text below

I heard only deepseek's provider does working prefills, but havent tried it

[blazing folio]

yeah this is what i get

i have no idea

it auto picks openinterface and deepintra as the common providers for me

[lament geode]

deepseek/deepseek-chat-v3.1 is a constant battle now. We don't use free models. Account is always topped up. Suddenly: No endpoints found that support tool use. Is this no longer supporting tool use?

[digital snow]

@nocturne kelp https://api-docs.deepseek.com/updates#deepseek-v31-terminus

💆‍♂️

[vapid torrent]

Hopefully it actually fixed the language mixing nonsense

[nocturne kelp]

wwtf

[jolly lintel]

#1419676592082518039

[lament geode]

I still get a 404 from deepseek/deepseek-chat-v3.1 at OR when my payload defines tools: regardless of provider. What did I miss or is this terminus unrelated?

The models endpoint reports that it supports tools.

[vapid iris]

would like to report that V3.1 (free) from DeepInfra appears to have a (temporarily?) reduced maximum context size to 49299 tokens (it appears to not be a consistent limit, 48k was a good buffer), anything above that results in a 404 'no allowed providers are available' error. Have ignored OpenInference to check.

Would it be possible to check with DeepInfra if it's a temporary or permanent change, and if it's permanent, be reflected in the provider list? Thank you!

p.s. apologies for the ping, just noticed your status of being on vacation, my bad!

[runic pike]

https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp

[vapid torrent]

https://discord.com/channels/1091220969173028894/1422148517081649202

[languid vigil]

love it that deepinfra did something on their back end to cause the quality of v3.1 to nosedive and struggle with endless repetition

[vernal wadi]

Why does v3.1 this sometimes? It's so strange. I didn't changed anything.

What is this shit?

<@&1384697330254610442> can you help?

It's not too much tokens

I don't understand what is happening

@storm seal any idea? My roleplay worked fine until now 🙁

[storm seal]

Hmm

Try Chutes as the provider

[vernal wadi]

How can I switch it?

I use my OR key

https://openrouter.ai/deepseek/deepseek-chat-v3.1:free

Only deep infra offers free

Then will I switch to 3.2 it's cheaper

The same issue O_O

Found the issue it's my prefill

[storm seal]

Ah, yayyyy

[vernal wadi]

But strange, it didn't happened earlier.

[storm seal]

strug

[void brook]

Some provider use different quant and also different stack
Mostly it direct you to provider that have problematic stack and weight

[vernal wadi]

What is a stack?

[storm seal]

The software they use to provide the service

Some providers might use vLLM while others use Llama.cpp or their own custom vLLM version

[vernal wadi]

Alright, that's out of my universe

[analog violet]

DeepInfra is permanently removing deepseek v3.1 or its temporary?

[storm seal]

Permanent

[analog violet]

Amazing

[languid vigil]

thanks janitorai

[vernal wadi]

Looks like v3.1:free now only works if you let them use your prompts :/

3.1 still works but only when selling your soul

[vernal wadi]

Yeah janitor had a huge impact on Open router, deepseek and chutes.
It's crazy h6huge the community got and how satisfied they all are from deepseek.
But I also don't know any alternative LLM which writes as good (and uncensored!) like deepseek.

And their built-in premium version of deepseek is with 20$ per month too expensive.

I also use chub 🫣😅

V3 is at <10% uptime. And v3.1 gets dead now

[marsh pulsar]

[storm seal]

Oh what the hell

[marsh pulsar]

yeah a 100+ billion difference in tokens is absolutely diabolical

[mild eagle]

Holy

[analog violet]

Its over guys...Deepseek free 3.1 is officially dead

No more free gooning

Only premium gooning for now

[coarse fable]

Since it died tragically, I have to admit I have never experienced so coherent and complex RP capacities with a 4-compressed model before 🥲

[surreal parcel]

So the free one is dead for good?

[storm seal]

Ooooh, it went to zero!

[vapid torrent]

yep

[vapid torrent]

also openinference doesn't allow gooning requests so no gooning to that so yeah

[coarse fable]

I got an idea that could kinda work for the free Provider in a "never happening hypothetical return(?)", Rate limits at specific services. The idea is simple, Free Janitor Users using a free model that got too many requests, they all get rate limit, like a single request per minute? Or two? Just affecting a single service instead of they all. That could not directly affect the other free users that used DeepSeek V3.1 for any other reasons, from helper to coding. It's kinda unfair that they all get affected because of Janitor users.

[gaunt edge]

Wouldn’t work - not good idea to punish third party apps for putting the app name in the api call. The obvious solution would be to just update it to not mention that it’s janitor ai, and then the situation is just worse

[coarse fable]

The hypothetical janitor ai updates are too smart—Anyways, seriously, Janitor would be capable of making its name incognito in api requests? Have they done that before or...? Damn, now that I think about it Openrouter can even get request from terminals, so they could not even do anything about a whole site going incognito.

[vapid iris]

they have to specifically put the name in the header to be identified in the first place

thats why janitor wasn't on any of the charts before late september

even though their usage number was already insane

[coarse fable]

I think here the problem is a middle between Janitor being hidden from view until last month and Openrouter for, well, just being an open router. Since it ends up affecting providers and users with situations like this, despite not being paid, the provider knew they would get anything, not even training data from "charity work"—But the may problem is whole provider being overwhelmed from a single site... welp, there is nothing users can do now, except either paying for requests... or simply getting subscriptions from other sites since Openrouter does not have subscriptions.

Honestly, Openrouter having subscriptions for certain models and providers would be an idea? like, certain amounts of dollars for a certain amount of daily requests so it isn't an abuse(?)

[vapid iris]

the only problem is with free models

since if they're paying, that's their own money to waste lol

[vapid torrent]

why did it be hidden? did they only recently add the title headers?

[vapid iris]

it was very recent, let me check when

sept 23/24

[coarse fable]

Yeah, why was it hidden?

So Openrouter can't track it well or simply they were dumb?

[vapid iris]

they didn't include it earlier, you have to specifically go out of your way to label it

it's opt-in not opt-out lol

https://openrouter.ai/docs/app-attribution

By including simple headers in your requests, your app can appear in our leaderboards and gain insights into your model usage patterns.
but it's more infamous than anything else 👀

[coarse fable]

Yeahh, anyways. If anyone here that loved to use "gooner rp free models" I think the Z-Ai is good, if you use a simple yet good system prompt with no jailbreak(because it makes reasoning too bulky)

[analog violet]

Off topic but using any free models with chutes as a provider feels like a gambling addiction

[coarse fable]

Yeahh, also feel it, but since Chutes added compression to their models they started to get quirky

[tough elk]

Won’t really use Z-AI, I’m more of a prompt person which is why i used deepseek and now it’s gone, forever probably

[coarse fable]

The big reasinoning Qwen of Vecine got a normal prompt mode if you put /no_think in the post history instructions, and with the recommended parameters of Temperature=0.7, TopP=0.8, TopK=20, and MinP=0. It seems to work

With a 10K tokens RP chat it provides good replies, just need to specify reply lenghts or else it gets messy long. And the only problem is the too many requests error

[strong moss]

Hey y'all, is the other v3.1 provider safe to use? Read here that they use your data to train their models. But isn't it the same for deepseek official?

What's

The best course of action here if you want to continue roleplaying with deepseek

[tough elk]

Thank and I also tried out GLM again and it’s the wildest model I’ve ever seen before. It gets straight into the not safe to work stuff and it doesn’t really matter what kind of chatbot your talking to. idk if it’s the model or if it’s my prompts since I have quite a list of prompts that are for the AI and definitely isn’t SFW so maybe it’s the prompt, I’m not sure.

[tough elk]

Unfortunately there is no way of role playing with the unfiltered version of Deepseek v3.1. Open inference does offer free Deepseek v3.1 but it’s filtered and you’ll get a lot of 429 errors similar to the other free Deepseek models on openrouter. There is also this one provider that I know but it’s also filtered unfortunately.

[coarse fable]

I used the GLM but honestly gets so many format errors that I genuinely give up xd

[signal sinew]

it's the model, even as an AI assistant it knowingly flounts ethical guidelines with minimal prompting

[lost carbon]

I use longcat.

😭

[opaque shore]

Wait can’t we just use deepseek v3.1 (not free) from openrouter? If we’re paying there shouldn’t be an issue right?

[mortal igloo]

Paid DeepSeek is better in terms of availability (100% uptime)

Yeah

[opaque shore]

Oh okay, I thought that was down too🥹

[mortal igloo]

I use directly from DeepSeek (no longer v3.1 but v3.2 Exp) and been having a blast since. Topped up $15 in March and still have $4 now

[opaque shore]

Is it less expensive to get it directly from deepseek?

[mortal igloo]

Low context (32k) and cache hits are a blessing

[glad basin]

deepseek is cheapest provider especially due to caching

[mortal igloo]

^In my opinion yeah because of caching

[glad basin]

you can enforce routing to deepseek on openrouter

[wild sparrow]

would you be willing to share your settings so i dont waste cash?

[mortal igloo]

I use SillyTavern (though all platforms can too), set context limit to 32k, and set Temperature to 1.2 (DeepSeek official API/OpenRouter provider converts this by deducting 0.7 so it's actually 0.5 temperature)

Now about caching, it saves cash because your previous messages are kept in their system (for hours in my experience), so if your earlier messages or system prompt changes, or there is a world info/lorebook entry activation (that inserts a new system message before the chat or within the chat), then the cache busts

If earlier context are kept and nothing changes from the very beginning to the latest message, then the cache stays.

(Except once you reach 32k of context: older messages are removed for new messages to be inserted, thus no cache and you get billed normally)

I don't know if I can explain better than this jumbled mess lol

[strong moss]

Would you recommend topping up to deepseek directly or on open router instead? I heard deepseek official is also a provider there so it works too right?

[mortal igloo]

Yep. You can just use the DeepSeek provider in OpenRouter
Though they only serve their latest model, which is DeepSeek V3.2 Exp

And the caching works too

I use direct Deepseek because Openrouter's deepseek account balance can be zero (no auto top-up) so errors haha

[brazen gale]

how do u do it??

[mortal igloo]

Choosing the official Deepseek provider, you mean? In SillyTavern you can choose a provider, but for other platforms that do not have a provider selection tab, I think you can just set "Allowed Providers" in your OpenRouter account settings to DeepSeek only (use the model DeepSeek v3.2 Exp, they only serve that currently and paid)

[vapid torrent]

You can also make a preset in openrouter

[brazen gale]

dang im all so new to ts

[wild sparrow]

how do you cache?

[vapid torrent]

Automatic

if you use DeepSeek provider

or DeepSeek directly

[wild sparrow]

like this?

[mortal igloo]

Yes

The caching process is automatic

[wild sparrow]

32k is a good amount for context as well

so that should last a while

[vapid torrent]

Be warned caching goes out the window after you hit ctx limit

[wild sparrow]

il drop 5$ in and see how it goes

[vapid torrent]

as the messages leave FIFO style

[wild sparrow]

how will i know?

[vapid torrent]

it’s cheaper to have long ctx compared to small ctx with DeepSeek’s 10x cache discount

[wild sparrow]

could you show your preset settings?

[vapid torrent]

I don’t use

i just set to maximum

so it never evicts from ctx window

[wild sparrow]

like that?

or whatever the max is for the model?

in this case 163840

[mortal igloo]

Yeah if you want full context, Deepseek official provider only supports 128k

[wild sparrow]

ive set it to 32k

so it saves money in the long term since i hardly go over 30 msgs

[vapid torrent]

My longest chat was 47k tokens iirc

[strong moss]

How does caching work? Does cache stay or something, considering that it's basically giving the bot an idea of the past events/roleplay?

It would work even if I use deepseek official through openrouter right

[mortal igloo]

DeepSeek keeps caches for hours / they don't say it in their docs. And not exactly past events per se, like cache is keeping whatever are the older messages are in the front of the context you've set. The bot will know what messages are in the context, not the entire chat history

[mortal igloo]

yes

[vapid iris]

according to their docs at https://api-docs.deepseek.com/guides/kv_cache,

Cache construction takes seconds. Once the cache is no longer in use, it will be automatically cleared, usually within a few hours to a few days.
so might be up to days, if their servers aren't under heavy usage.

[strong moss]

That sounds neat, but I wish it lasted longer since my writing could take hours too lmao

[vapid iris]

I saw people mentioning they had cache hit at 25 hours

[strong moss]

ah neat

[vapid iris]

but if you're doing a chat and your old messages get pushed out

might not be a cache hit anymore

[strong moss]

Since I roleplay using proxy, I often just use the same character (with the same definition) so the character definition would be consistent, would that mean it's always a cache hit even if it's a different chat but the same character?

heck, would it be possible to maintain cache for weeks, maybe even months?

[mortal igloo]

ohh never knew that. didn't read enough myself...

[vapid iris]

zero chance 😂

[storm seal]

You'd have to send 1 message every so often depending on the cache time

[mortal igloo]

yeah. no cache, deepseek will re-cache everything once you've gone past your context limit since the messages after the character definition (which would be the bot's first message) is replaced with whatever nearest message to fill the front context

[strong moss]

OHHH so the cache would only hit until something like... 16k if I set it so?

Would setting the context window to 32k/16k be cheaper than the 64k/128k context with cache

[mortal igloo]

Yep, if you set context to 16k, it'd cache the 16k like:

1. System prompt [500 tokens]
2. Bot definition [1000 tokens]
3. Persona definition [100 tokens]
4. Chat history [1st bot "greeting" message, 2nd message your reply, 3rd bot reply, 4th your reply, etc. totaling 15400 tokens]

This is what the 16k cached would be.

Then if you send your reply to the bot once the context is full, the bot's 1st "greeting" message would be replaced by your 2nd message from the chat history like:

1. System prompt [500 tokens] <-- this is cached since nothing changed
2. Bot definition [1000 tokens] <-- cached
3. Persona definition [100 tokens] <-- cached
4. Chat history [2nd message your reply, 3rd bot reply, 4th your reply, 5th bot message, 6th your reply, etc. totaling 15400 tokens] <-- no cache since the bot's greeting message is replaced by your reply // there is a change. DeepSeek re-caches this repeatedly if you still use 16k context

So if you decide to use 128k, the cache would prove useful until you hit 128k. More than that, it re-caches the whole 127,4k tokens

I use 32k context because when it re-caches, deepseek bills me $0.01 instead of $0.03 if 128k

cheap

[crimson egret]

Okay, guys. Can someone explain to me whether V3.1 is still working or if it’s completely no longer free? Some people say it can still be used, but one of the endpoints was removed, while others say it no longer works and has been shut down.

[vapid iris]

paid v3.1 is working just fine

v3.1 free had two providers, DeepInfra and OpenInference

DeepInfra no longer provides free v3.1 due to the massive human horde from janitor

and OpenInference is a far smaller provider

so you're just going to get 429 errors

[crimson egret]

So now it can’t really work properly on the free model?

[vapid iris]

if you thanos snap 95% of the janitor horde, it'll work. but currently you're competing with way too many people for the available capacity.

[crimson egret]

Thanks for the reply.

[paper magnet]

rlly? I was under the impression 3.1 free got axed then taken out the back and executed because of the graph shown earlier that it was over. I got 3.1 free from OpenRouter and I keep getting errors whenever I try to use it, and when I go on the error link, it says its no longer available

[vapid iris]

Its effectively dead because of the number of users

[willow junco]

If my context size is at 32k now, does that mean once I hit 32k context, it'll be cheaper?

[mortal igloo]

there'll be caching when you build up the context/the chat gets longer until 32k, and as I mentioned normal 32k input price is $0.009 or 1 cent
so yes, cheap

[willow junco]

🤔 I see. Thanks! I was a little confused since it costs $0.009 for me. ^^

[mortal igloo]

oh wait

my calculation's wrong, sorry. It's $0.009 or 1 cent...

damn

[chrome rune]

Hello, I need help. I get this message when I try to send a message in any chat on “Janitor AI.”

[rocky sedge]

3.1 is dead

[chrome rune]

Shit

[hushed hedge]

you can still use it from platform deepseek

[abstract igloo]

you can still use it in OpenRouter, it's just the free endpoint that's gone

[hidden plinth]

Does anyone else have a problem with DeepSeek V3.1 putting non-english characters in tool calls? It consistently does it for me, and causes tool calls to fail.

[tropic kettle]

do you mean https://openrouter.ai/api/v1/chat/completions is gone and that is the reason for these response?

{"error":{"message":"Not Found","code":404}}

[abstract igloo]

no, i meant there isn't a free DeepSeek v3.1 anymore

[abstract igloo]

also, depending of what you're using, you should only use the https://openrouter.ai/api/v1 part

[tropic kettle]

i am just trying to test the endpoint jst as it is in the documentation and i am receiving 404

const url = "https://openrouter.ai/api/v1/chat/completions";
const headers = {
"Authorization": Bearer ${process.env.OPENROUTER_API_KEY},
"Content-Type": "application/json"
};
const payload = {
"model": "deepseek/deepseek-chat-v3.1",
"messages": [
{
"role": "system",
"content": "hello"
},
{
"role": "user",
"content": "If you built the world's tallest skyscraper, what would you name it?"
}
],
"temperature": 0.7
};

const response = await fetch(url, {
method: "POST",
headers,
body: JSON.stringify(payload)
});

const data = await response.json();
console.log(data);

any response?

[abstract igloo]

idk, it worked here

try seeing if you have any provider preferences in your settings

[jovial idol]

how do i see the reasoning tokens for v3.1 and v3.1-terminus

here is my code. i tried a lot but cannot see the reasoning for different kind of user messages.

Any help would be much appreciated.

client.chat.completions.create(
model="deepseek/deepseek-v3.1-terminus",
messages=[{"role":"system","content":system_prompt},{"role":"user","content":user_message}],
temperature=0.1,
top_p=0.1,
stream=True,
max_tokens=2500,
extra_body={
"provider": {
"sort": "throughput"
},
"enable_thinking": True,
"reasoning": {
"effort": "high",
"exclude": False
}
}
)

[storm seal]

can you pose the raw, full response here?

[abstract igloo]

i think it needs to be
"reasoning": {
"enable": true
}

[lost carbon]

if you include the effort param , reasoning is enabled by default

[sweet zodiac]

i can't find the good things form v3.1,
i use R1 and V3 in the past and the prompt(s) fit them.
I believe the DS family trained by similar dataset and behave the same.
however whne i change the model from R1 to V3.1, the result ruins.... 😂

[spice marsh]

Different models need different prompts, even in the same lineup

[lost carbon]

biggest learning tbh , hence you have to be mindful when you build agents.

(DSPy fixes this)

[dense wolf]

I dont see Deepseek as a provider for 3.2 exp though

And setting all other providers for the model to ignored just gives me a no endpoints error

[abstract igloo]

you probably have to enable this in your settings

DeepSeek

[dense wolf]

I did enable it just now and still got the error

Let me reopen my app though

Just in case

Seems to have worked

[atomic quartz]

Hey folks, does anyone know why OpenRouter is throwing an error?

[storm seal]

You disabled allowing providers to log/train on models

[humble echo]

in a chubapp i see deepseek 3.1 base, 3.1 terminus and 3.1 chat

What is the difference (?)

[spice marsh]

3.1 Base -> Won't chat with you, it will complete your sentences
3.1 Chat -> It'll be able to chat
3.1 Terminus -> This is 3.1 Chat with minor bug fixes (some repetition issues and mixing english and chinese)

[humble echo]

thanks bro!!

[atomic quartz]

Hmm, where and what should I turn on?

[storm seal]

turn on the first 3 options in the privacy section of your account settings

[atomic quartz]

Oh, thanks. But after doing this, the chat shows an error: "No allowed providers are available for the selected model." Thanks for helping

[storm seal]

Did you black list Chutes?

[atomic quartz]

No, I don't even know that there is some kind of list )

[vapid iris]

Do you have option 5 (ZDR) activated?
It should be turned off.

[atomic quartz]

I didn't enable this option.

[storm seal]

I'm not sure why you can't use DeepSeek v3.1 (free)

[vapid iris]

can you show a picture of your allowed/ignored providers section at https://openrouter.ai/settings/preferences ?

[atomic quartz]

Is this the providers section at ?

[vapid iris]

yes, can you show the top too?

[atomic quartz]

Oh, sorry, the top point didn't make it into the screenshot above.

[vapid iris]

Hmm, this is weird. All your settings are correct, V3.1 (free) should be working fine for you, hmm.

[atomic quartz]

The account isn't banned. ) Z.AI: GLM 4.5 Air (free) also doesn't work

1. Could the models not working be related to geolocation?
2. Is there technical support for Openrouter?

[vapid iris]

I'm not sure on that front, sorry. Maybe someone else will know, but I'm not aware of OpenInference geo-blocking.

[atomic quartz]

Is this a problem?

[abstract igloo]

i don't think it should be

[buoyant willow]

anyone else's v3.1 free model just being weird

it js gives "1.1..1.1.1..1.11" or similar gibberish

[storm seal]

Try another provider

If using a different provider stops that, then you know it's that particular providers fault

And you all can report that provider here

[buoyant willow]

Theres only 1 provider for free rho ;-;

OpenInfdrence

Ill se

[storm seal]

Ah

Oops, I didn't know that

[buoyant willow]

I tried both yesterday and the day before, im not sure whats the issue

Via oenrouter chatroom

and aider

[storm seal]

Set your temp to 0.3, and if it still happens then it's confirmed its an Open inference issue

See if you can reproduce it in a new chat if it's an Open inference issue, and Toven will yell at OpenInference to fix it

[buoyant willow]

Ah so it seems it defaultd at 1 for xhatroom and 0 for aider, ill try with 0.3

One sec js gotta get back home

[buoyant willow]

Appears 0.3 works but for some reason it gave me an end of sentence token at the end which appeared on the chat message? https://openrouter.ai/chat?room=orc-1760793523-IhYneTn3lo9GCKT4SW3N

But aside from that it works 👍

[storm seal]

I can't see the chat btw

Chats are stored locally, not on server

[buoyant willow]

Oh wait hih

huh

Nvm its spewing nonsense again

It may be a specific token count cuz it only happened when i pasted a large block of text

[storm seal]

Can you post that here?

[void brook]

Base on my experiences if we use the non-free providers and chose the good provider even with large block of context i don't get that

Mostly because of the provider of free models
Which understandable for free models, but i guess if they can fix it then it's amazing of them as a provider

[mortal igloo]

Perhaps because of the quantization? OpenInference provides free DS v3.1 at int8 precision which is lower uner fp4 I think

[gaunt edge]

int8 is less quantised than fp4 (so, int8 is higher quality). for some reason i thought openinference v3.1 free said fp4, but i checked and it does say int8

[void brook]

Agree on this, but i don't know with those new NVFP4, i heard if it goes through the pre-train, sft then rl phase with NVFP4 format it will be comparable with the BF16.

I guess that make sense, because deepseek when they doing original training with FP8 the model actually have comparable performance with BF16 than if it come from BF16 then being quantize to be FP8.

The key point is in what format does the model originally being train on.

[unkempt stratus]

Yall I'm using the og 3.1 chat

Not the terminus

And it's lagging so hard, I'm using paid

What's going on

[full relic]

i am trying to make a local language ai chat bot using the deepseek v3.1 from openrouter but the url to get the request seems to be broken i am using https://openrouter.ai/api/v1/chat/completions for api but the chatbot shows url not found what could be possible cause and a solution to this?

[full relic]

Seems to work now

[unkempt gorge]

response tokens are coming through empty and every thing is ending up in the reasoning field today for some reason? tried across several providers all of which were fine before

[vernal wadi]

It loaded endless long to finally respond with a very short message. But consumed 3x the amount of tokens then my input was?
I assume it has "invisible" thinking?

Not sure if that is good or bad...

[abstract igloo]

if you click to see more info you can see how many of them were reasoning tokens