Deepseek V4

[flat osprey]

since gemma tends to be more token efficient

[opaque reef]

deepseek reasons in character

for roleplay

[pastel sail]

yes thats right it would cost about 6.5x more based on artificial analysis data

also its gonna be way slower because of this

[hot swan]

that's why gemma is way on the cheapest side of that line yeah:

[woeful jay]

flash seems really solid

wish they wouldve added multimodal

yeah it seems that pro is kinda underwhelming for what it costs

at least its not benchmaxxed

[sharp vortex]

Tbf v4 is supposed to release last month, but it got delayed in their technical note

[simple mauve]

I saw this somewhere, too. But the model name doesn't say preview. So what's the deal? Is it a preview, or is it not? 😮

[green trellis]

For coding?

[sharp vortex]

https://api-docs.deepseek.com/news/news260424

They did say Preview

[simple mauve]

Yeah, but the preview models are (usually?) named ...-preview, and this one isn't. I'm confused.

[sharp vortex]

I personally think fully releasing model will be 4.1 with engram. So they don’t need to make preview name

Or just that they don’t satisfied with current v4 yet

[jovial kelp]

Google also doing this thing, i don't remember since when google use non-preview for gemini pro

[sharp vortex]

Wdym Gemini will never out of Preview stage
-# they will release new preview model before GA old model frfr

[abstract dragon]

Why is nobody hosting the model

Is openrouter no longer worth it anymore due to the high traffic it provides?

[hot swan]

if it's getting high traffic it's clearly worth it, isn't it

[gusty sphinx]

hope it was worth it "io-net"

[simple mauve]

Probably because it's a brand new model that came out on a Friday, and it's now Saturday which is not a normal workday, and it takes time to properly set up a new model.

[raven canyon]

its still working fine for me

i think its probably just got high load

[simple mauve]

yeah, they went straight to my ignore-list

[abstract dragon]

No lol you're wrong

First of all this is a hugely anticipated release. Second of all a bunch of providers have set it up already but they're not serving on openrouter for this specific model.

And thirdly it absolutely does not take time to deploy this. Any person with a bit of gpus can set it up within a few hours. Let alone big inference providers

[gusty sphinx]

i don't know man, it used to be that providers would happily serve their busted ass implementations that were technically "working"

which sometimes lead to some really poor publicity for a model, because most aren't paying attention to who is providing it

or are even aware of how this works

[raven canyon]

deepseek v4 has new implementation details i think

[abstract dragon]

Idk I'm just starting to feel that openrouter aint worth it. Almost all models experience low TPS and im assuming the openrouter accounts on these inference providers have their own quota. And I'd be better off just registering directly in those providers to get better tps

[gusty sphinx]

it wasn't that long ago that most providers found out what model "tool calling" is

[raven canyon]

i think openrouter counts TPS weirdly

[abstract dragon]

Like im so done with 20-30TPS models and most of open source models on openrouter are experiencing that

[abstract dragon]

Bro the models feel slow af

[gusty sphinx]

i mean go find out right now and report back

maybe i'm wrong

[abstract dragon]

I dont have to

Take cloudflare for example. When k2.6 came out the openrouter cloudflare provider almost died. While cloudflare's official api was stable

[raven canyon]

then use BYOK i guess?

[gusty sphinx]

ok. i didn't try k2.6 for a while

[abstract dragon]

I will, just expressing my frustration

[gusty sphinx]

cloudflare is good

i'm just saying, if you think this is a shitshow, it used to be a lot worse

[abstract dragon]

I know because i used open source models on openrouter a year ago

But this isnt an excuse for slow models. One model can't call tools but it responds another runs at 9tps and 20s latency

[potent lagoon]

Is togetherai still busted?

[oak maple]

yea just saw that

holy moly 75% off :o

[obsidian walrus]

trust the plan, deepseek always delivers

[faint belfry]

I get random numbers in sentences from Pros responses.

[potent lagoon]

yesss, it's on togetherai provider

[faint belfry]

Is there a way to force it to only go through Deepseek provider on JanitorAI?

[brisk sand]

make a custom https://openrouter.ai/docs/guides/features/presets

[hot swan]

holy hell

[faint belfry]

Thank you sir

[hot swan]

I thought the price decrease would be like 30%, not divide by 4

I think the deepseek team actually fucked up massively not delaying release by a day

if those were the prices they were about to offer

imagine the narrative and the cost/output graphs if those were the prices for pro at launch

incomprehensible blunder

[cloud flame]

https://tenor.com/view/sensual-sentinel-gif-5787468

[raven canyon]

yeah cause at that price it’s way more good value

only 3x the price of flash instead of 12x

[thin bramble]

damn, now the price to performance makes sense.

[covert topaz]

call me schizo but did they update the model

😭

[raven canyon]

deepseek v3 level pricing for … who knows? level of intelligence

[indigo folio]

🙌

[covert topaz]

1m whale slaves

[indigo folio]

filling oceans

[rain shuttle]

Lolll expert and instant think so differently

[short jasper]

They updated what

[covert topaz]

idk feels different

[sharp vortex]

WE ARE SO BACK

[short jasper]

i told you bro

they would change

smth

Since it says

[covert topaz]

but they didnt announce anything

[short jasper]

preview

version

[short jasper]

yeah they dont announce anything

smart market

they collect feedback and fix

efficient

[covert topaz]

in 1 day?

its the weekend

[rain shuttle]

The same expert in Chinese, I guess it does some bias based on languages also

[potent lagoon]

nice more providers

[sharp vortex]

they wait until benchmark period is over then discount fr

[covert topaz]

not different in a bad way btw in a good way

[short jasper]

let's feed deepseek more data

[covert topaz]

thats why i was thinking of changing my prompts to chinese

[sharp vortex]

Do gooner really care that? (it's anonymous data)

[covert topaz]

might get better output

[sharp vortex]

-# pretty sure no one actually coding confidential data with Deepseek anyways

[short jasper]

so you change your prompt to chinese and it follow instruction good?

[covert topaz]

nah it would be dumb to code w deepseek

kimi and glm are way ahead

[covert topaz]

its what ive heard not confirmed but people did it in the past and claimed to get better responses with older deepseek models

the problem is getting it to output in chinese

[sharp vortex]

Proof or placebo effect

[covert topaz]

i mean english 💀

[hot swan]

kimi is certainly better but at the current discount? I could use deepseek

[covert topaz]

https://tenor.com/view/proof-pro-proo-cat-laptop-gif-22485700

[sharp vortex]

ain't no way Deepseek place a high price trap so they can "discount" when other provider start to host it

[hot swan]

look at that insane score/price ratio for flash

meanwhile deepseek pro is currently just twice the price while scoring 71-87 (depending on thinking budget)

[sharp vortex]

tbf flash might be good model if it's multimodal fr

it's suck when it's text only

[simple mauve]

How does that work across different providers, though? One provider has version X under the same name as another provider's version Z?

[hot swan]

it's not even clear if kimi is better at coding in fact

[dusty birch]

how do you use max in api? xhigh reasoning effort?

[raven canyon]

i think

[sharp vortex]

yeah

[dusty birch]

alr

ty

[sharp vortex]

what site is that

[hot swan]

benchlm.ai

[raven canyon]

what makes it better than any other benchmark site or Artificial Analsyis

[hot swan]

🤷‍♂️ it's pretty?

I mean I like that they have a ton of ways of comparing

[raven canyon]

is it a benchmark itself or an aggregator

[hot swan]

aggregator

[elfin sparrow]

true

[woeful jay]

this is COMPLETELY wrong btw

[abstract dragon]

https://gif.fxtwitter.com/tweet_video/HGJgCprXQAAjU21.webp

[woeful jay]

where is mimo flash

[hot swan]

it's considered free so not ranked as such

it scores ~61 though

[dusty birch]

i cant get a single request through Pro, just ratelimits

[pastel sail]

id try siliconflow for inference

they have it properly setup

decently more expensive than the official api though

[lime moth]

flash versión is bad for roleplay

[hot swan]

define "bad"

[pastel sail]

far sloppier than pro

less world knowledge and dumber

[lime moth]

dont follow the format, random thinks, change the narrative from third perspective to 2 persona, etc.

alot of rerolls, cheaper but dumb

[lime moth]

i will try the pro

[hot swan]

it's surprisingly competent at fiction writing in general though

[covert topaz]

use pro for rp

[thin bramble]

that is the fault of the greedy providers

[haughty pilot]

deepseek today?

[crude steppe]

just plain evil

they make the providers look greedy

[covert topaz]

Chads

[cloud flame]

Considering cache, you can cut of actual prices even more

[gusty sphinx]

https://orca.orb.town/monitor?model=deepseek/deepseek-v4-pro guess whos back

(its io-net)

[cloud flame]

Shady's back

[wild mango]

Deepseek provider is having issues with v4 pro? I can only use it via other providers

[hot swan]

and they've still got the worst service on offer

[fresh edge]

Does 429 Provider returned error really mean over-load?

A lot of endpoints doesn't support tools? Or OR is missing something?

[gusty sphinx]

also https://orca.orb.town/?q=deepseek+v4

[covert topaz]

https://tenor.com/view/black-guy-laughing-burst-laugh-burst-out-laughing-can't-hold-back-laughter-hold-back-laughter-gif-8627946458514606624

[simple mauve]

are they just trying to cash in on a new model not having too many providers yet?

[covert topaz]

it has a fair few now and theyre still trying to steal through fallback routing 💀

[twilit lodge]

guys, kinda new to AI pricing and terms. The input/output tokens... thing.
lets say I chat ~200 messages a day (roleplay chat). If I used Deepseek V4, how much I'd be paying each day for these 200 messages?

Is there a way I can estimate that?

[rustic island]

Well, a million tokens are, depending on the tokenizer, around 4 million characters

You would need to calculate how many characters in + (out + reasoning) you use and do the math around that

But there's also caching, repeated input tokens (so e.g. re-sending the chat) will be discounted depending on the provider

But I think I'd just simulate a typical long chat with your usage for some turns and see how that acales

[twilit lodge]

Hmm, guess the best course is to put a few bucks and see how long it lasts, then. But for deepseek do you put up credits in DS itself, or in OR?

[supple sigil]

with how bad the rate limits are for OR right now, probably just in DS itself

i think most apps support the DS api natively so you should be able to just skip OR altogether

[rustic island]

There's BYOK for that

It's your choice, really, both should work fine (other than perhaps the limits)

Not aware how the OR limits are right now

[twilit lodge]

rate limits? like restrictions?
again, not familiar with AI terms

all I did was set up OR account, use one key for free models in the past. So yea, not much experience

[exotic elk]

I keep getting a 500 Internal Service Error.

[wild mango]

best course it's to put a few credits and keep an eye in the logs, and keep in mind that the longer your chat = the more expensive each request gets. Also if you're interested in NSFW chatting, you might want to exclude a few providers that don't allow it

[twilit lodge]

aight, thanks for the info
gonna take a look into it

[marsh goblet]

Is Gemma 4 better than DeepSeek Pro V4?

[stray pulsar]

How good is it, based on initial impressions?

[supple sigil]

extra latency and cost past 1m requests

[rustic island]

Uh lol

Will you really ever send a million requests a month for RP?

I don't think the extra latency makes a difference either

[supple sigil]

probably not but i still think its better general advice

if whatever youre using already supports the deepseek api, then why route through another api

[simple mauve]

Isn't that more expensive, though? DS seems to have an extra 6% VAT added on the topups.

[supple sigil]

probably, but at least there’s the benefit of avoiding ratelimits

[sharp vortex]

Doesn’t OR get almost 10%

It’s 8%

[simple mauve]

VAT? Not necessarily. I'm vat-exempt, for example, yet would still need to pay DS the 6% extra. At least I found no quick way to specify that it should be 0%

[exotic elk]

Still keep getting Internal Server Error. I have ZDS, but that shouldn't apply to SiliconFlow I don't think.

[sharp vortex]

I see

[feral scaffold]

just block every provider except direct api

[vapid karma]

Unfortunately direct provider is getting clawed to death at the moment

[feral scaffold]

openclawed to death

[exotic elk]

Nice. Got it to work on my text adventure via Together. Good results and cheap too.

[lime moth]

Even in pro RP, have errors like dialogue ( -Like this- ) and the prompt say dialogue ( "like this" ). Everytime give me dialogue with that -idk why-

[woeful jay]

https://tenor.com/view/please7tv-please-beg-pray-hope-gif-3860103425950934673

[exotic elk]

I absolutely hate it when it gives me a reply where all the text is in the Reasoning. It basically stole my money.

[tulip estuary]

it did it many times for me

[jovial kelp]

Which provider and what frontend?

[exotic elk]

Together, and I'm using OR chat.

[jovial kelp]

Try different providers first, tell me how the result for different providers

[exotic elk]

Rn V4 is like an old lawn mower engine. Have to keep pulling until a request gets past the 500 error.

[jovial kelp]

Yeah, OR really need to get some deal to increase their rate limit for each providers.
I remember someone talking about how they got rate limited in OR but when they go straight to the providers site and use their service directly, it allow them to make more request than in OR.

[exotic elk]

io.net is "provider ignored by account." No it's not. Only AtlasCloud is on my banned provider list.

Evidently I can't use the providers Deepseek and io.net, though I removed the ZDS on my Guardrails. SiliconFlow I can use but I keep getting an error 404 (No endpoints found for deepseek/deepseek-v4-pro.). So Together is the only game in town, and 9 times out of 10 it doesn't work either.

[jovial kelp]

Yeah, it's a bit hard with deepseek if we don't want our data to be use for training

Because deepseek it self is still the best serving provider for their own models

[sharp vigil]

I'm confused as to why a provider would be banned for using users data for trainin

[simple mauve]

It's probably best to wait a few days. It's not surprising that with zero workdays passed since the launch, not all providers are at their best.

[sharp vigil]

I thought that's what everyone did

And it was a given

[simple mauve]

No. Most providers respect your privacy.

[jovial kelp]

Some people just doesn't want companies to use their data for training because it could contain private information

Option is good in the market imo, if you don't have problem of giving your information then go on but if you don't want that then there should be option for it too.

[exotic elk]

Just to experiment I switched off "Always Enforce ZDR" and saved. Deepseek still isn't a provider.

[simple mauve]

You also need to allow data trainings. And make sure it's allowed in the default workspace as well.

Using workspaces is still a little confusing... I just recently realized that it doesn't matter if I create a new workspace where I want to allow Deepseek as a provider--if it's not allowed in the default workspace, it still won't show up in the new workspace.

[raven canyon]

its cause there are two levels, account and workspace

what you should do if you want to do that is:

• allow training account level
• disable training on default workspace
• allow training on second workspace

[simple mauve]

Oh... oh... so what I thought was the default workspace, is not even the default workspace, but the account-level setting, meaning I will have to copy every setting I've used so far over to the default workspace. Oh boy... Okay, well, at least it's doable, even if quite a bit of extra work. Thanks for this bit of info--this was a missing piece of the puzzle I didn't even know was missing.

[raven canyon]

the account level settings should be also in the default workspace

its just if you want to change something

[simple mauve]

Yeah, I get it now. So far I've had DeepSeek (the provider) blocked, and that's still how I want it in the default. I wanted to create a new workspace specifically for DS today, but couldn't. Now I understand why. I need to copy over the account settings into the default workspace, then change the account settings, then create a new workspace for DS specifically. Yeah, doable, but not tonight, LOL. 😄

[odd badge]

I noticed that too. He was doing it naturally, without me prompting him in any way. I thought it might be a issue or hallucination. Maybe it's a training thing?

[plucky ermine]

Are they literally the only paid provider that trains on API data?

Kind of wild

[potent lagoon]

All of them do

[plucky ermine]

If you want to go conspiracy mode and assume that every single provider breaks the law, sure, but that clearly isn't what I'm talking about

[exotic elk]

Are any of the other providers going to go down in price? Deepseek is conspicuously cheaper.

[raven canyon]

deepseek flash has some providers but it looks like they don't support tool calls
deepseek pro is also looking not great

[crude steppe]

https://tenor.com/view/speed-kill-that-boy-speed-kill-that-boy-ishowspeed-kys-gif-10749022239239578206

[raven canyon]

deepseek flash options: (all same price)

• DeepSeek offical: may train on your requests
• DeepInfra: has tools, but 7tps is rough (potentially reporting error?)
• SiliconFlow & NovitaAI: decent uptime and speed, but no tool calling

deepseek pro options:

• DeepSeek offical: cheapest by far, may train on your requests
• GMICloud: second cheapest, but not great uptime, and no tool calling
• SiliconFlow: second most expensive, decent uptime, no tool calling
• IO.NET: most expensive by far, fast (potentially), unknown uptime, only non-deepseek with tool calling

so basically do not use openrouter yet for deepseek pro requests with tool calls cause if deepseek official is rate limited your request will go straight to IO.NET with 5x-10x the cost

[frank wind]

pro is very creative in text completion with no instruct formatting

low probabilities but never incoherent

1.1 temp, no other sampling

honestly I am surprised

[worldly pier]

It's a gold rush and laws historically haven't mattered that much in gold rushes. If they earn more selling your prompts and completions for training than they do providing inference, then of course they're going to do that.

[pure flax]

it's main thing is the knowledge base since its large. Hopefully later versions make it smarter / less crazy

[frank wind]

yes, indeed. it knows a lot

[jovial kelp]

Don't forget about stability and quality, third-party inference providers are good options but in some cases and it's a lot, the performance of the models get butcherd

[plucky ermine]

Assuming an unethical actor, it isn't selling prompts vs selling inference, it's selling prompts vs the cost of a lawsuit for breach of contract and permanently being shamed and blacklisted out of existence.

Regardless, it matters what they promise. I'm not going to a restaurant with a terrible health inspector rating just because you go "eh, they're all dirty, they just hide it." Okay, well this one definitely is, so I'll take my chances somewhere else.

[gusty sphinx]

https://tenor.com/view/fat-guy-shooting-gun-gun-shot-gif-15114243

[thin bramble]

wtf, i am doubting myself now

https://tenor.com/view/fuwamoco-fuwawa-mococo-フワワ-モココ-gif-9169317763107150779

[gusty sphinx]

you may need to update your prompting strategy sir

[thin bramble]

claude biased due to claude synth data in deepseek?

EQ-Bench 3 is a LLM-judged test judged by Claude Opus 4.6, evaluating active emotional intelligence abilities, understanding, insight, empathy, and interpersonal skills.

Longform Creative Writing Benchmark: Judge upgraded: Evaluation now uses Claude Sonnet 4.6 (replacing Sonnet 4).

[thin bramble]

i guess

https://tenor.com/view/cat-cats-pet-cat-cat-pet-cute-cat-gif-24810247

[gusty sphinx]

i think it's generally agreed that claude is biased towards itself. but i've used gpt-5.4 quite a bit and "emotional intelligence" is not the first thing that comes to mind

[worldly pier]

v4Pro uses a fuckton of "it's not X it's Y"

[gusty sphinx]

i haven't dug into the benchmark strategy in a long time, but it probably is more "accurate" than what most people would see from a small amount of casual use

[thin bramble]

leme see arena again today

still low votes

[elfin sparrow]

yup

[sharp vortex]

It’s not sonnet anymore? 💔

[thin bramble]

it is sonnet for long form.

[sharp vortex]

Opus is too expensive for long form ig

[plucky ermine]

What I've found on the main EQBench 3 is that the ELO isn't reliable, but the category scores usually are

[thin bramble]

also short form creative bench means nothing

[woeful jay]

no dont i swear 4.7 is a regression too

[raven canyon]

good release or underwhelming release?

[woeful jay]

pro underwhelming

flash good for price

[flat osprey]

pro isn't super underwhelming - just not a good deal for the price

right now with the discount it's actually great value

[loud verge]

In terms of Artifical analysis scores. The deepseek v4 flash(max) is 3x more efficient price to performance wise than the model that gets second rank

All in all, deepseek didn't disappoint

[flat osprey]

yep, lines up with my pareto graph

I think it's a really good option for creative writing, but I honestly prefer Minimax M2.7 or Gemma 4 31B for cost effective coding

mainly because the hallucination rate on V4 flash is absurd and it also just thinks for quite a long time

[loud verge]

Yeah, at max thinking it thinks too much. But it's still really good at high thinking and thinks a lot less (about 1/3rd of what it does at max)

but even the high thinking module thinks more than the previous 3.2

[flat osprey]

yeah, it's also hard to determine how much intelligence is lost when you drop from max to high

[woeful jay]

yea i agree, i phrased my thing wrong

[loud verge]

yup

[pure flax]

mimo 2.5 pro is the real unsung hero

[loud verge]

These models are still preview. Deepseek probably gonna drop a banger soon

[flat osprey]

yeah mimo 2.5 pro is really good honestly

just a bit pricey for my liking, but really capable

[sharp vortex]

Deepseek tmrw again 💔

[flat osprey]

i also think V4 was unfortunately severely overhyped - like, people thought this was gonna be opus-level reasoning at the price of V3.2

like, without the hype, V4 pro is still amazing value for what you get - it's just that better options exist now

[loud verge]

I mean it's deepseek. They did bring r1 which was the performance of O1 at the time while being like 4% of the cost.

[flat osprey]

which is fair

they mainly got there because LLM research was still in its infancy though - a lot of the tricks they used for R1 just don't apply anymore for the same gains

[cloud flame]

What are the tasks that suffer much more and much less from high hallucination rate?

[loud verge]

It brings miracles. That's why I worship the whale 🐳

[flat osprey]

mainly questions that have conflicting or outdated training data

like KP's notorious test of asking how to fix lag in a Paper Minecraft server

[loud verge]

P much everything matters a lot. But high stake things matter more than low stake ones.
So if you're researching, writing code and maybe something medical, then hallucination matters a lot

[flat osprey]

sometimes the model will think it knows something well enough that it just won't research

[cloud flame]

So the more tasks needs to calculate smth vs the using existing knowledge, the more hallucinations affect it

[flat osprey]

yeah pretty much

[cloud flame]

Yeah, I wouldn't use this model for medical lol

[flat osprey]

like, hallucination rate becomes a lot less relevant if you're always relying on the model researching

but it still is a factor

[cloud flame]

Fun fact: before Gemini 3.1 release, older Gemini 3 previews also had high hallucination rates on all models - both Pro and Flash. The way they combated it is advising all developers and users of Gemini 3 to use 'web_grounding' web search native tool call, which worked basically as "ARE YOU SURE ABOUT THAT?" John Cena behind model's back during thinking process

Unfortunately, Deepseek v4 does not have that exact option

[plucky ermine]

Yeah Gemini had terrible hallucination rates (I think Flash is still god-awful?) and was really bad on safety stuff too like enabling mental illness.

[flat osprey]

yeah, you usually have to supplement V4 with a research tool like OR's research option or a custom-built one

for instance - i regularly use Gemma 4 31B with research, and even though the model has a high hallucination rate, i can get outputs comparable to non-thinking frontier models

[flat osprey]

yeah flash hallucinations are still pretty bad lol

gemini is just mentally unstable overall

[plucky ermine]

Honestly the negatives don't get enough attention IMO. Hallucinations and mental-health matter a lot, especially in consumer apps.

[flat osprey]

yep, which is why 4o had (and arguably still has) a cult following

hallucinations + sycophancy are a match made in the psych ward

[plucky ermine]

I know to use Opus to call me out on my bullshit, but the average person doesn't know how any of this works.

"Arguably"? They still post on Twitter all the time =P

[flat osprey]

true lol

[thin bramble]

yes.

[cloud flame]

So deepseek has native search?

[flat osprey]

speaking of opus - 4.7 feels a lot worse at the role of calling out bs

sometimes it just calls out things that are either fine or weren't an issue to begin with

[flat osprey]

no model really has native search

it's always a tool call or tool process

[plucky ermine]

It's also subtle. The model specifically needs to be skeptical and blunt imo, because when people vent they don't give the other side of things. So it's easy for even good models to take your side when it shouldn't.

[cloud flame]

I mean native provided

[flat osprey]

oh, then yeah it does lol

don't think they provide it through the API though like Google does with grounding

[plucky ermine]

I need to test that more with 4.7

[cloud flame]

I can build a local web search with injecting data into prompt, but it won't be as good as smth provider of model does on their side

[flat osprey]

it's not even like it's sycophantic - it just misses key facts really easily sometimes and jumps to conclusions too quickly

[flat osprey]

i would look into Tavily. they have a pretty generous API for fetching web data naturally for LLMs

it's what i use for my custom research tool

[plucky ermine]

I was amazed that GLM-5.1 called me out to just the right degree on something a while ago in a test. Better than anything else. Was sympathetic but critical, clear but not assholish, etc.

[plucky ermine]

Reasoning on or off?

[cloud flame]

Isn't any 3rd party web search tool just do another round trip with sending input context? Or does it activate after user sends prompt, before model processes it and decides to call tool?

[gusty sphinx]

it's actually great for medical roleplay

[cloud flame]

Like what?

[cloud flame]

https://tenor.com/view/outlast-outlast-groom-darling-scary-gif-14024561

[flat osprey]

on, but you'll find that on claude web it's a bit difficult to get it to think deeply on something

the model very frequently just... doesn't research stuff it should be researching

[regal shuttle]

Anyone know whether there are representatives from DeepSeek on this Discord?
I saw from a post on X that they are asking for feedback, but their own Discord appears to be an unmoderated, malware filled mess.

[plucky ermine]

Basically hit it with a concern about one of my friends. I kind of stream of consciousness my thoughts in and then run it across all the good models in OR chat.

It had deeper insights than anything else, covering observations and ideas that other models skipped past. It correctly called out spots where I might have biases, am being overconfident, etc.

[flat osprey]

none that i am aware of

they would've likely said something here by now if they were around

[plucky ermine]

Aside from being genuinely useful as a way to sort through my thoughts, I really like it as a model test. They have to figure out what I actually want, since it's largely venting. And what I need but don't want. Meta observations, like taking into account that I chatted about it. And the problem itself of course.

[regal shuttle]

A shame. Would be more than happy to give feedback etc, but not taking out an X account to bother, and their own discord appears to be a waste of space.

[raven canyon]

i also vouch for Tavily, i use it, the free tier is good value

[raven canyon]

is that with or without the discount

[heavy sable]

Have anyone noticed that NovitaAI, SiliconFlow not working? its only DeepInfra working when excluding Deepseek own infra

[indigo folio]

i knew this would happen

it's so good

also kimi droppedddd on longform writing, damn

[indigo folio]

it's very delicious

the slop going down by times 2 compared to all their previous versions

[flat osprey]

without

[covert topaz]

holy

[covert topaz]

how come?

ain’t no way there’s model degradation already 💀

[indigo folio]

yea...

10.5 points down

compared to 2.5

[covert topaz]

[flat osprey]

probably has to do with how heavily 2.6 is tuned for logic and coding

[indigo folio]

yea

[covert topaz]

i love me some agentic workflows 😝

[indigo folio]

their creative writing elo is up by 300 points though

not bad

[covert topaz]

well idk if it’s just me but I noticed alot of claudisms popping up more with DS

[indigo folio]

i'm sure they're distilling from opus too

[covert topaz]

the thing im trying to avoid

[indigo folio]

i expect their creative writing elo to be up too

it's at almost 1600 rn

[covert topaz]

yeah more than likely

lol GLM and DS competing over who can distill the best

[indigo folio]

bring it on

[cloud flame]

Kimi K2.6 falling in longform is very weird - should be other way around

[indigo folio]

the only people it hurts are the billionaires

[indigo folio]

yea i thought it'd be better than 2.5

[covert topaz]

well people who are sick of claude too

[indigo folio]

ds team does test and experiment a lot

claude distills are just a bonus, but they don't really depend on them

[covert topaz]

they acknowledged rp feedback on X btw was kinda cool they want English speaker feedback

[flat osprey]

imo any open models before the bubble pops are great to have
development will slow down rapidly once VC money runs out

[covert topaz]

so could get some rp tuning later down the line

[indigo folio]

yeaa this isn't even the production version of v4

just a preview

so they're gathering feedback for the official v4 production version

[covert topaz]

It was similar to 3.0 preview from what I recall

[indigo folio]

i do think they're building the current base so they can add a lot of features and improvements without big problems

which is amazing

[covert topaz]

that model sucked at instruction following (for me) then 3.1 it religiously followed them

[indigo folio]

and this model still runs on the old gpus

which is crazy

[covert topaz]

can only get better as more chips come through

[indigo folio]

yep

[elfin sparrow]

fr

[indigo folio]

expect the performance to be much better with the new gpus

[covert topaz]

DS got a bright future fingers crossed

[indigo folio]

and a huge drop in price

for pro

[covert topaz]

anthropic.. falling off

[indigo folio]

these people must be panicking tbh

[covert topaz]

got beaten by oai

[indigo folio]

the os are closing in on them

they have to be much better to be an exception now

[elfin sparrow]

not anytime soon

[covert topaz]

mm doubt, unless they do a 180 after opus 4.7

[cloud flame]

4.1 tomorrow

[indigo folio]

and honestly, the ds team shows that intelligent and skilled people matter, a lot

not just the money

[elfin sparrow]

XAI kinda falling off

[indigo folio]

look at what they did with the old gpus

[covert topaz]

were they ever good tho

[elfin sparrow]

what?

[indigo folio]

the current v4 runs on old gpus

[elfin sparrow]

at one point ig

[flat osprey]

they had their 2 day leads on artificial analysis before lol

[indigo folio]

that's why it's a bit more expensive

[vale kayak]

recommend tempature is 1.3?

[elfin sparrow]

aren't they running on huawei chips?

[indigo folio]

once the new gpus arrive, the performance will be much better and the prices for pro will drop by a lot

[covert topaz]

saw that earlier tried it, sucks for me

[cloud flame]

V4 thinking ignores temperature

[covert topaz]

low temp better for me

Imagine thinking

[indigo folio]

1.5 for cw

[elfin sparrow]

even if they were running on old GPU's how come GLM is dominating?

[indigo folio]

has anyone tried that out yet 😭

[covert topaz]

yes

it’s ass for me

[indigo folio]

both great companies, but glm's prices are up by a lot

so they're barely scrapping by tbh

the only os models i really recommend are kimi, glm and ds

i don't trust anything else

[covert topaz]

I was really impressed with GLM in 5 then they just went the coding route with 5.1

and it’s definitely cause of openclaw blowing up in China they wanted to take full advantage of that

[elfin sparrow]

yea

[indigo folio]

the real companies are the ones not sacrificing cw

for coding

[covert topaz]

think they’ve all done it atp

[flat osprey]

mimo v2.5 pro is very good for its price

[indigo folio]

https://x.com/ChineseEmbinUS/status/2048281093921538411

Chinese Embassy in US (@ChineseEmbinUS)

#DeepSeek #AI #china #ChinaTech ✨

X · Apr 26, 2026, 6:00 AM

fr

[covert topaz]

✨

[indigo folio]

i've seen people say this, so probably. granted, i've never tried a lot of other os models besides the 3 i named

[elfin sparrow]

gemma 4 31b on par with ds 4 flash

[indigo folio]

and qwen

[covert topaz]

yes I keep hearing praise about this one

[indigo folio]

gemma 4 is good too

i use it for my discord bot

[covert topaz]

agreed

[indigo folio]

its personality and sassiness got me tbh

i love it

[covert topaz]

if it had the smarts it would’ve been godly

[flat osprey]

i mean it does, just not much world knowledge

[indigo folio]

what i really ab glm is that

[flat osprey]

if you give it research it gets very close to non-thinking frontier models

[indigo folio]

it actually writes you a long novel

if you ask it to

[covert topaz]

ya for its size it’s way ahead but I mean like SOTA level

[indigo folio]

i told it to write me a long novel

and it gave me a 15k word one

[covert topaz]

https://tenor.com/view/fire-writing-gif-24533171

[indigo folio]

waiting to go above 8-9k words

[covert topaz]

did u read it

[indigo folio]

with other models

not the whole thing, but i did read a little and even got claude to analyze it

said it was pretty good

https://x.com/HarshithLucky3/status/2048346214983508036

Harshith (@HarshithLucky3)

DeepSeek V4 Pro can generate "up to" 384k output tokens

gave it a prompt to test it

Its been over 10 mins and its still writing

Lets see how many lines of code this thing drops and how it looks

Will update when it finishes

X · Apr 26, 2026, 10:19 AM

this is crazy

https://x.com/HarshithLucky3/status/2048394352372957521

Harshith (@HarshithLucky3)

After 30 mins it finally gave me:

5317 lines of HTML code
~100k tokens

This is just on the DeepSeek official site

Don't look at the UI. I didn't mention anything about the UI in the prompt

I will try with another prompt to push it to full 345k tokens output

▶ Play video

X · Apr 26, 2026, 1:30 PM

[gusty cradle]

Deepseek updated price again more cheaper!!!

[short jasper]

so did deepseek change their reasoning style again?

[indigo folio]

the goats

[hot swan]

once again I ask why they didn't just launch with those prices

could have made it clear it's a limited time offer and launched with those discounts anyway

[oak maple]

wait wtfff, that's really good. also seems like its not a temporary change?

[sharp hamlet]

New age for gooners

[elfin sparrow]

dammn

[hot swan]

this is the reason we're constantly hitting the rate limits btw

[cloud flame]

https://tenor.com/view/cable-guy-jim-carrey-future-future-is-now-gif-16404601

[covert topaz]

ignore me I thought I was using non thinking version but it was thinking so my parameters were disabled, high temp way better

https://tenor.com/view/ipad-kid-playing-minecraft-and-watching-youtube-shorts-ipad-kid-thingstoryx-gif-4913695962410316975

[cloud flame]

You can send 2 or 5 as temp with thinking - it won't change, yeah

[short jasper]

so does this mean non thinking deepseek v4 pro is recommend for roleplay?

hmmmmmmmmmmmmmmmmmm

[covert topaz]

noticed it with 2.0 temp as an experiment I was confused why it was still coherent lol

[covert topaz]

not sure tbh

I think so

being able to control my parameters makes it soooo much better for writing

it’s weird cause having a role enabled set as assistant at the bottom disables the thinking process but ur still technically on thinking mode

[wild mango]

Deepseek is out of control

[rustic island]

Wait what

No word on this being temporary? 👀

That is essentially free cached input to me

[wild mango]

I think it's permanent considering they put the actual pro price reduction, from 0.145 to 0.0145

[hot swan]

yeah, 0.003625 is the 75% promotion on top of 0.0145 <- 0.145

and unlike the 75% promotion the 90% cache discount doesn't have any message about a date limit so

[wild mango]

the promo got the reduction too, because it was originally 0.036 🤣

[rustic island]

Crazy

$2.8 per billion input tokens lol

[hot swan]

indeed (for flash)

[cloud flame]

Flash with that pricing and cache could be used for something big and static running on cron every 1-2 hours and be super cheap to maintain, even with reasoning

[hoary zenith]

what the absolute fuck

[mighty arrow]

Wonder how high that puts V4 on the efficiency curve, that caching change in particular

[rustic island]

Like someone said, what a waste of PR lol

If they had launched with this 75% discount and that cache discount, I guarantee that'd have made many headlines

[hoary zenith]

The way they move, I think they are still traumatised by R1 PR

[covert topaz]

I think they originally thought the pricing justified the model performance then did a last minute switch up after the pricing being negatively received

[pearl geode]

So how good is this

[cloud flame]

Samantha REALLY didn't like it

[vivid tide]

yeah. i dont at least not for RP

but everyone has their preferences

[cloud flame]

I don't

[vivid tide]

aight

[simple mauve]

Not for RP, but which one? This has so many versions now.

[short jasper]

deepseek v4pro is good for roleplay, Just thinking mode loop output ruins it completely..

[vivid tide]

v4 pro. i dont enjoy it. both with and without reasoning. my own thinking formats also the pretrained roleplay ones. in either user prompt or system. its not good to me

[covert topaz]

it seems like it needs a lot of wrangling but once it gets going its really good I’ve been tinkering with it for a few days and now it’s at a place where I prefer it over my go to model (gemini 3.1) u need to account for its lack of instruction following (strong reinforcement injections etc)

[vivid tide]

but im not here to convince anyone on that. it doesn't fit my preferences for creative writing, even after heavily altering presets for it. that's it that's all. its a decent assistant chat bot tho.

[covert topaz]

had to make alot of adjustments to my preset all around and my personal cot is basically new lol

[sage kraken]

what's the recommended way to use v4 without thinking on sites that don't have a specific toggle for it?

[thin bramble]

same here, shit model.

https://tenor.com/view/ruby-copium-oshi-no-ko-推しの子-ruby-hoshino-gif-4151121735588077996

[supple sigil]

you cant

disregard, i was wrong ^

[thin bramble]

or set thinking to minimal if that is an option

[rustic island]

You can, even if the site doesn't have a toggle
Just create a preset (Preferences > Presets) that forces no thinking

[thin bramble]

guessing prefilling but ya

[short jasper]

I think we forgot mimo 2.5's existence

let's compare mimo with deepseek, If mimo is superior

📢

[cloud flame]

With vanilla pricing maybe, but not with that discount

Soon they gonna pay for people using Deepseek

I want to findom a LLM provider

[rustic island]

Maybe a dumb question given DS's reputation, but has anyone ever gotten banned from the DS provider? I'm sort of concerned about deploying this and people pressing the model to speak about Chinese politics

Their ToS are surprisingly strict and forbid NSFW (lol)

[supple sigil]

i wouldnt worry about it

if openrouter hasn’t been banned yet, you probably won’t either

[covert topaz]

based on what ive heard people say about it for rp its probably better but im happy with the way i got ds setup now

[wild mango]

I'm getting "Provider returned error" in 9 out of 10 requests when using v4 pro, it's so annoying

[grizzled spear]

I just discovered that v4 is out, now I checked the uptime and... I will not even try it.
It's far away from 90%

The orange is most important because I (and others) usually use other providers and they are available now at under 40%!

[tulip estuary]

i'm using directly through their API

i'm not hitting any limits

[grizzled spear]

Is it there better?

[tulip estuary]

yes for this case

launch week is crowded

and we're basically sharing one API key from OR

[grizzled spear]

The issue is not a limit, the issue is that through the platform openrouter the access is just barely able because the connection is not very stable, probably is there a too weak server

[tulip estuary]

the issue is the rate limit

the Provider returned Error

[grizzled spear]

Hmm okay

[wild mango]

yeah, Cairo is right

I decided to put a few bucks in their API directly and haven't had any issues so far, so It's on OR side

[simple mauve]

The DeepSeek provider has an insane cache time... and with these cache prices... that's seriously impressive. I've just tested the cache after an hour, and it's still there... documentation says the cache stays from a few hours to a few days... The only issue of course is that the data can be used for training... but with these cache conditions, it's really worth considering. Most other providers have a cache time of maybe a few minutes.

[tulip estuary]

it's because they don't have a huge limit from the provider (DeepSeek) like they have with other major labs

[wild mango]

it's weird because it only affects pro, if I switch to flash I get no errors

[tulip estuary]

because Flash is less rate limited, it's a much smaller model

[wild mango]

oh so the limit is by model not by provider

[tulip estuary]

it's both

[wild mango]

maybe OR sets a higher limit in the next few days then

but now I have some credits on deepseek api so I'll burn through those before trying OR again

[tulip estuary]

"burn" isn't even a correct word for this absurd pricing

i'm eyedropping through my credits

it's not very good for coding though, it's fine

gets a lot mixed up

[soft fulcrum]

Flash?

[cloud flame]

I got full cache hits after ~19 hours of sending previous prompt

[tulip estuary]

pro

[feral scaffold]

Hope it comes to the tavo mobile app soon

[indigo folio]

the goats 😂

deepseek using an untraceable distillation technique

[mellow jewel]

Time to boost Dipsy with my gooning sessions.

[indigo folio]

🙏

[raven canyon]

im using deepseek BYOK openrouter

[pearl geode]

Okay so not the best for rp but what are benchmarks

[pearl geode]

That looks pretty good

[indigo folio]

yep, the longform writing's slop is at least 2x lower

than ever before

[knotty root]

ngl feels kinda mid

[pastel sail]

creative writing is kinda of a meme bench

it has very strange placements

like pony alpha is glm 5 and its above glm 5, glm 5 is also above glm 5.1

The creative writing leaderboard is being updated to use Claude Sonnet 4 as judge (previously used Sonnet-3.7). The top models have already been updated; the remainder are a work in progress.

it literally uses claude sonnet 4 for rating.

[plucky ermine]

It is useful, not perfect. The real RP benchmark to me is Sillytavern usage

[tame swallow]

Judged by Opus 4.6

oh

gdi I make fun of myself once more

you are right

it's sonnet 4

but why is this weird method of judgement?

one bench they be using opus 4.6

one did used sonnet 4.6

and this was on Sonnet 4

[glass cradle]

anyone succesfully using deepseek v4 on vscode+github copilot?

[meager kelp]

pony alpha was an alpha of glm 5, not the same model

and as I keep saying, undertrained models are better at writing

[woeful jay]

i kinda doubt that tbh

the turnaround from pony -> glm 5 was like

instant lol

idt they did any training in that time period

[meager kelp]

there'd be no point in releasing a seprate alpha if they did no training

and it's also very clearly different subjectively and in several benches

[raven canyon]

they’re probably doing some for a post training while the alpha period was running

or maybe they just did that to collect prompts

[woeful jay]

yes

[woeful jay]

0% chance

they were hammered for compute during the stealth model period

[covert topaz]

I LOVE DEEPSEEK V4, thank yuo for your attention to this matter

[pastel sail]

no way

training doesnt take that little time

[pastel sail]

0 to 20 chance that it was a diff version of glm 5

it was extremely similar and they released within days

[pure flax]

deepseek v4 pro is better with a long preset. Its dumb and needs to be told how to write but its knows a ton and is super creative. I actually like it now
Now it reminds me of old opus, opus 3 maybe

[pastel bluff]

Has anyone actually managed to get DS V4 pro working?

All I get is rate limits on all three providers except the overpriced io.net

[sharp vortex]

You either need to use direct api or use other providers 🥀

[pastel bluff]

Classic

[covert topaz]

exactly

literally the same perspective i have lol

it did really well on my creativity test

[simple mauve]

"Insufficient Balance" ehh... OR ran out of credits with DeepSeek 🙁

Just when I wanted to give this thing a good test, LOL. 😄

Yeah, time to bite the bullet and open a direct deepseek account to get an API key... 😄

[peak swallow]

Direct api is so much better. Just add it as BYOK in openrouter

[sacred glade]

how so?

[peak swallow]

No rate limits and low latency

[simple mauve]

Yeah, it just costs slightly more. But probably still worth it if I don't run into the global ratelimits.

[simple mauve]

Well, actually it's only 1% more than going with OR, so yeah, that's completely negligible.

[acoustic hawk]

One message removed from a suspended account.

[sharp vortex]

BYOK doesn’t cost anything for a long time now

It used to be 1% before

[simple mauve]

Yeah, but OR adds a 5.5% fee on topups, DeepSeek adds a 6% VAT. I thought OR added just 5%, but it's 5.5, so the difference is only an extra half a percent with DeepSeek directly.

[sharp vortex]

Or you somehow get a way to use alipay (it got no tax iirc)

[steep summit]

Hmmm will try this next time

[sharp vortex]

It look funny when they don’t want to combine currency lmao

[west shell]

how does Deepseek manage to be like 4x cheaper than the second cheapest provider? Is it because they do prompt logging? Or is it a marketing thing?

[simple mauve]

They most likely want some good training data. That's just my personal speculation.

[west shell]

that kinda what I suspect too, but I suppose I cant be sure

[simple mauve]

And keep in mind that the pro pricing is just a temporary discount.

[west shell]

I see

[gilded onyx]

do you use flash or pro for RP? what reasoning level?

[pastel sail]

most providers are overpriced and its actually a high margin business, its just training models spends 10x money

deepseek is probably providing the model at a minor loss right now but the original price was reasonable actually

[sharp vortex]

Deepseek only host 2 model at the time too

[simple mauve]

Flash is better with reasoning disabled?

[cloud flame]

It's weird, maybe mixed up

All other benchmarks and every model I saw on UGI always show reasoning enabled improve scores, or in worst case stay the same

[glass cradle]

anyone succesfully using deepseek v4 on vscode+github copilot?

[vapid karma]

I always take it with a grain of salt, but Livebench is looking pretty good too

Also lines up pretty close with the statement that V4 Pro is around ~GPT-5.2, intelligence wise

[lime moth]

they literally put the LLM free for anyone, they really dont care about money.
Cheap prices, for useful data training.

I love this Chinese LLM.

[west shell]

but like

its not sustainable

I love free lunch as much as the next person, but it literally cant last

even if you get "infinite, free" training data

[lime moth]

Free for those who can set up their own server. Low prices for those of us using the model from a paid server; nobody sets an unviable price to make their business fail. If it's priced that way, it's because it's barely profitable; I suppose the other providers are just very ambitious.

[meager kelp]

Servers are not free to run

[vapid karma]

They also said the discount on Pro input / output was temporary (through 5/5 I believe) so it'll probably be back to the higher pricing afterward

Until they get their node set up, anyway

[jovial kelp]

Deepseek is not the CEO primary source of income, it's like side project for him
Seems to working more closer with the china goverment, they got more in home built gpu being use for inference so it should be cheaper because it's in home built and chinese special discount with other chinese company

[twilit lodge]

Guys, never tried getting paid credits for DS
You make a api key on DS site, put up credits there, then use BYOK on openrouter, and thats all? Do you need to config anything else? (token limit, something like that)

[jovial kelp]

Their strategy have always being providing inference in breakeven category, not getting as much profit as possible

[simple mauve]

If you've used OpenRouter before, then what you've said is pretty much it. You can top up with as little as $2 on DS. You may want to set OR up so you ONLY use that API key and provider, but it's always good to have fallbacks if DS is down for any reason, so you probably don't want to mess around with that either.

[west shell]

I see

[feral scaffold]

It doesn't spaz out when you crank up frequency penalty, for pro at least

[cloud flame]

Because it probably ignores it

Thinking mode just ignores all samplers possible, even temperature

[timber obsidian]

Hi guys anyone know what the best providers are for Deepseek V4 flash? Novita is the fastest but doesn't mention the quantization so does that mean it's a nerfed version?

Maybe I should use the DeepSeek platform then? Do they have US servers

[dusty birch]

im finding pro pretty good at coding, it picks some pretty good choices for changes, it doesnt overengineer very much

[opaque rapids]

extended to end of May

[sharp vigil]

We love deepseek 🐳

[toxic rose]

Endless extension

[frank wind]

winrar type shit

[feral scaffold]

they need to train on those inputs

[opaque rapids]

to me there are 2 possibly case,

1. being bait for them to collect the data and further RL
2. burning the money until the Huawei chip getting properly deployed

[frank wind]

train on my smut please

[opaque rapids]

I guess now I don't have to worry running 10 multi layered agent to write my smut novel

[sharp vigil]

Agent swarms are actually possible now

[gaunt dirge]

Say thank you and feed them your Opus prompts.

https://tenor.com/oFPjNTwMoxv.gif

[jovial kelp]

W distiler

[minor canopy]

Guys, why Deepseek v4 pro/flash doesn't work on OR ? I've tried with opencode and Kilo code but can't use it. Keep getting errors from "provider" [AkashML] deepseek/deepseek-v4-flash is temporarily rate-limited upstream. Pleas...

[onyx bramble]

try different providers or make yourself a dedicated deepseek api key on their site, rate limiting on OR deepseek is pretty rampant right now

[minor canopy]

Well, I'll do this

thank you

[gilded onyx]

When they say non thinking mode it's just setting reasoning to none, right?

[raven canyon]

yeah

[gilded onyx]

Anyone using DS for roleplaying? Would you recommand flash or pro? At what reasoning level?

[jovial kelp]

For creativity no reasoning will be the way to go, but if we still want some hard consistent logic in it then use the highest setting possible.

I am using their chatapp so the option only thinking and non-thinking

[gilded onyx]

I'm using their api directly. I'm building a roleplay app for romance and nsfw. It's kinda hard to say which settings is better

[blazing bough]

hi, when will fireworks be available? (with tool call support)

[exotic elk]

I'm using Pro and XHigh reasoning. Good results.

[cloud flame]

For me sending xhigh (which maps to Max) gives almost same results as sending high (which maps to High), maybe it should be as that?

[exotic elk]

So Deepseek's pricing is a temporary discount through May?

[gilded onyx]

i'll give it a try again

[gilded onyx]

yeah it seems like it

[exotic elk]

Not sure how much I like that. The thing is, at the non-discount price, it's not much cheaper than Sonnet, so I may as well use that. Or just go back to 3.2.

Almost as good is only really a sell when it's significantly cheaper.

[indigo folio]

they'll lower the prices after the new gpus arrive

on top of this until-may-31-discount

[onyx swift]

Anyone know why novita doesn't have tool calling listed as supported on either flash or pro?

They have other models that do show support, and their own model page suggests it should be supported (looks like novita calls it "function calling")

[tulip estuary]

i'm yet to see a 15 second long reasoning

[wild mango]

Discount on pro now extended until the end of may

[exotic elk]

How does the cached thing work? My 50k context text adventure turns are bafflingly cheap. Like $0.0014 a turn, whereas it should be around $0.02 a pop.

[tulip estuary]

see the image above

everything you send to them is cached for about 24h

so when you send it back with a new prompt, only the new prompt is counted as input cache miss

plus the output tokens price

[wild mango]

At these prices we are basically paying for output

[exotic elk]

Ah, ok thx. And when are they getting these new GPUs that'll bring down the price?

[tulip estuary]

they said second semester

[rugged vigil]

yay, (almost) free DeepSeek V4 \o/

[jovial kelp]

Holy

This model is really good if you wanna learn unconvential knowledge

Finally, good less align model

[simple mauve]

The flash version is just plain stupid even with max reasoning. It's funny in its own way, but definitely not something I would use in a production environment.

[covert topaz]

[covert topaz]

i heard complaints about it leaning towards being positive but I myself think it’s been neutral so far less unhinged than gemini tho

[thin bramble]

trick/jb gemini instead

[jovial kelp]

I don't know about that
Giving it the same prompt as what i give to Kimi K2.6, DeepSeek-v4 giving me the answer but Kimi K2.6 rejecting it.

It being to positive could be true, if it keep on trying to satisfied user by answering even bad query then it technically showing a positive behaviour toward the user.

Similar case with claude model who want to be overly helpful that it accept bad query and completing it.

[dusty birch]

a image understanding beta is on the website

as mentioned in the paper they will work on multimodal models

[cloud flame]

Vision mode probably would be small separate, instead of on top of Pro

[sharp vortex]

Or v4 flash has vision 🗣️

[raven canyon]

is instant v4 flash and expert v4 pro?

[sharp vortex]

Yes

[toxic rose]

Finally, vision-ds

[gusty sphinx]

what sort of unconventional knowledge?

exotic knowledge? dark knowledge?

[toxic rose]

anti-knowledge

[rustic island]

Lol, this model is a softie

I ask it to be critical and analytical but it feels like it eventually feels bad and starts toning it down after ~15 turns
Not quite in a sycophantic way, it just softens the tone

[gusty sphinx]

who could ever stay mad at you

[regal shuttle]

It really is; have noticed exactly the same, not just in chat-bot mode but also creative-writing/roleplay. Not sycophantic, but has a real tendency not to play hardball. It can of course be forced to with prompting, but where choice exists it leans to the softer approach every time in my testing compared to e.g. GLM 4.6 (GLM 5 also a softie), or to Gemma.

[sharp vortex]

Whale plushie take over

[cloud flame]

https://tenor.com/view/orca-orcane-orcane-plush-orcane-woo-gif-25280880

[dusty birch]

theres some interesting vision/multimodal tokens in the tokenizer of v4 pro and flash
also some box, polygon, point and ref grounding-like tokens

these tokens dont seem to be present in 3.2 tokenizer, nor OCR 2 nor Janus

these tokens when given randomly seem to trick the vision model from the website to hallucinate random stuff

[left grotto]

deepseek v4 pro through openrouter (deepseek provider) repeats my system prompt to me in every response
is this a deepseek bug? or an openrouter bug

[exotic elk]

How does D4 compare to Sonnet 4.6?

[mellow jewel]

OR/provider bug

[broken lintel]

just because I'm curious will you get banned from deepseek as a provider through openrouter for any china related info that comes through as prompts?

I'm planning on having it go through some state legislature bills and some of them mention the PRC/Taiwan and I'm uncertain if I need to be filtering that out to other models

[simple mauve]

Depends on what you use it for, but they're different leagues.

[eternal thorn]

Kinda a beginner question but is there a way to get it to generate long reports on a topic like you can do with Gemini and claude natively or do I need to make a custom agent for that

[supple sigil]

every time a new oss model comes out you hear that it’s a “sonnet replacement” or “opus replacement”. those statements are usually complete bullshit. v4 pro, to me, actually is a sonnet replacement - borderline opus replacement. it is incredible at agentic tasks, reasoning, coding, etc. over the past 4ish days ive done ~50m tokens (only cost me like $2 lol) in agentic coding and deepseek genuinely is nearly flawless.

tl;dr i would only recommend using sonnet over deepseek if you prefer the style or like spending money

[tulip estuary]

i really dig the deepseek style

it was a bit sloppy before but much better than whatever GPT is always doing

[broken lintel]

I think everyones threshold is different tbh, for most things m2.7 was mostly fine, V4 is definitely a nice upgrade though and the price is just too good

[soft fulcrum]

What do you think of V4 flash?

[supple sigil]

haven’t tried it nearly as much as pro, mostly just because pro is already dirt cheap

considering v4 flash is larger than minimax m2.x series models iirc, and it’s a deepseek v4 model, ill assume it’s pretty good

[thin bramble]

or just continued pretraining like kimi k2.5

the "upgraded" model has vision

[feral scaffold]

I'm still tinkering around with prompts, but it kinda feels worse than v3.2 so far. Pro is a bit better

[broken lintel]

running cost estimates for deepseek is almost stupid ludicrous if the caching holds, my workload is very cacheable and just did an estimate for one run through. About ~20 mill uncached input tokens, ~421 million cached tokens, and about 110k output tokens would come out to about $4.0266 ? 😭

Has anyone ever had cache miss on things they shouldn't through openrouter or generally is openrouter good at not interfering with that?

[rustic island]

Did you pin your provider to DeepSeek?

[supple sigil]

i keep thinking "surely i have to put more in by now right?" and i keep not having to put more in. the value from deepseek right now is absolutely absurd 😭

[wild mango]

Anyone knows how long the cache lasts? I was hit today with a 1.5 million tokens cache miss 😭

[raven canyon]

1.5 million token? the model context isn’t that big

[wild mango]

Yeah I don't know what happened 🥲

[covert topaz]

idk maybe i was lucky on release but the writing felt fresh on release day compared to now, now all i see is we have claude at home, the same sentence structures isms, phones buzzing, low (insert sound here), fluorescent lights yada yada i just want to get away from claude man 😭 like i managed to prompt it out but its just a drawback that i picked up on

im attributing it to the giant ass discount on their pro model so everyone and their mom is jumping on it -> model degradation

[oak maple]

the prompt logging is a big downer though :/

[thin bramble]

i changed my mind on rp, on web, expert thinking deepseek isn't bad.

[lime moth]

Bros, anyone testing in RP NSFW???

Temperature and reasoning?? low, medium??
Plz help for long RP (with repetitive scenes).

[thin bramble]

(both in web, with thinking) dipsy v4 expert sometimes mogs glm 5 turbo in certain stuff

[thin bramble]

deepseek sometimes reads the room better, and glm lacks behind because it is too focused on the character card.

https://tenor.com/view/anime-coffee-poker-face-gif-16351805043609296883

[thin bramble]

with all models temp 1, reasoning max, helps with long context too in smaller models. if it is repetitive, then prompt issue ig.

[dusty birch]

seems to be related to this
https://github.com/deepseek-ai/Thinking-with-Visual-Primitives/blob/main/Thinking_with_Visual_Primitives.pdf

[dusty birch]

yup

I see a <|ref|>cat<|/ref|><|box|>[[x1,y1,x2,y2]]<|/box|>

yeah this is deepseek v4 flash

[soft fulcrum]

Is there any way to stop Deepseek from putting reasoning into the final answer?

[supple sigil]

maybe try increasing the reasoning effort if you don’t already have it set to max

[soft fulcrum]

does that actually fix it?

that's weird

I have it set to high

it's a shame that it needs max, since it's already so inefficient with high thinking

I will try it

it went from 2000 tokens of thinking to 13,500 just from high to max

this is ridiculously inefficient

[supple sigil]

my thought process was more reasoning before answering will cause less reasoning while answering

kind of just shifting where the reasoning happens

[soft fulcrum]

yeah

In theory, deepseek seems smart, cheap and fast. But in reality, it's slow (because of thinking), still cheap but way less than you'd expect, but probably still smart

I wish they focused more on efficiency this time

I hate how companies use reasoning as an excuse for laziness, at least that's what it seems like

[woeful jay]

not gpt

5.4 brought good token efficiency improvements and then 5.5 made it REALLY efficient w reasoning

(well yea they increased the price but its understandable tbh)

[hoary zenith]

looks like the whole repo was taken down, wonder why

[left grotto]

can confirm it stopped doing that when i went deepseek direct api

[mellow jewel]

Works every time.

[rich ferry]

Providers half-assing new model releases is one of the biggest pain points I see with OR

Unfortunate

[flat osprey]

yep, that and providers just sticking with absurd pricing from some companies like qwen

[jovial kelp]

I think that will never change, they need to put more effort to make multiple models work as how it being intended and most of them either just doesn't want to much effort into it or aren't actually able to do it

At the start of 2025 i also feel the same with a lot of providers

[sharp vortex]

Yeah Qwen put absurd price because they want ppl to use Qwen-plus or flash series on api instead.

But other providers no longer cutting price anymore 💔

[cedar tree]

cc @deft crow I know you already know, but just want to emphasize this as an ultra important trust thing. Honestly feel like exacto should be the default or something

[jovial kelp]

isn't exacto only for structure output?

[cedar tree]

yeah, I guess exacto but even more

i want openrouter to vouch for these providers are "as good" as the reference, basically

so if they're using vLLM main and vLLM main is borked? good luck, you're not getting served

i'd honestly love to see openrouter do the same thing that SemiAnalysis is doing and run SGL and vLLM reference models, and compare against 1p upstream + compare against other providers

[deft crow]

can you share api call details? haven't seen this from others

[deft crow]

very much top of mind for us

[rare gale]

For what it's worth, in a production capacity, I would be perfectly paying a decent surcharge just to get better reliability/consistency ; maybe that could help fund the persistent benchmarking/testing?

[gusty sphinx]

OpenRouter Exacto Pro MAX?

[deft crow]

funding isn't really the problem, running good benchmarks and using that signal correctly is extremely hard and time consuming

i've been trying to hire a like head of evals

[rare gale]

For sure, totally understand the challenge but I think even a health ping every few seconds would atleast better flag which providers are just straight up broken. Obviously easier said than done but just wanted to note that there's probably many others out there that will gladly pay a decent margin to ensure some degree of reliability.

[ocean night]

V4 flash was rad at coding typescript in a small test. It decided to use the write tool instead of edit repeatedly but never made any errors.

[raven canyon]

if you ask deepseek v4 "what is in the image?" it will hallucinate an image for some reason

note: only happens for non-thinking mode

[left grotto]

i just tried that and got "I don't see any image attached to your message. Could you upload the image you'd like me to identify? You can attach it directly to your next message and I'll take a look."

do you get this consistently? i wonder if its a bug in openwebui

[raven canyon]

i set reasoning.enabled = false

[left grotto]

in what api?

[raven canyon]

openrouter via deepseek byok

the openrouter chatroom seems to not be able to disable reasoning (?)

[left grotto]

oh the actual openrouter library, im using the openai library, not sure if it would make any difference

[raven canyon]

shouldn't

maybe its a quirk with deepseek's official API

they might be injecting something in the system prompt

[left grotto]

does deepseek have a library or do you mean their direct endpoint?

[raven canyon]

im using it through openrouter

deepseek provider

this is really inconsistent

[left grotto]

earlier i had it repeating the system prompt to me every message and deepseek direct api fixed it
might be worth trying to see
ill regen a few times and see what happens

[raven canyon]

i am effectively using deepseek official api

i have deepseek provider as BYOK in openrouter and have forced it to only route to it

[left grotto]

not necessarily, it still goes through all of openrouter's layers
openrouter was always doing BYOK but with their own key essentially

[raven canyon]

was this using deepseek provider or a different provider?

[left grotto]

that was using deepseek provider through openrouter
then i used deepseek.com endpoint directly and it stopped
today i tested again and openrouter seemed to have fixed it

[left grotto]

i think i may have deleted the conversation cant find it, but testing again today it seems like the problem is fixed
will let you know if it happens again

[raven canyon]

i just tested, using deepseek official API it has the same weird behaviour

[left grotto]

ya i regenned like 5 times each on openrouter, deepseek, flash, pro
it seems to realize i have no files are attached

could genuinely be an openwebui bug i had lots of trouble when i tried to use that before i just made my own chat ui instead

maybe try hitting the endpoint directly and see if the behavior is the same

like just copy out a python/curl blob from openrouter docs

[left grotto]

ah yep i see what you mean

[raven canyon]

what

[left grotto]

now deepseek is just flirting with you 😂

i just told deepseek to make an svg of a cat and i got back chinese talking about reptiles lol
never seen that before

[raven canyon]

"what is the image" => "i cannot see the image" 8/8 times
"what is in the image?" => random hallucination

[left grotto]

interesting

[sharp vortex]

Deepseek vision tmrw

[vapid slate]

Is this the first one?

[sharp vortex]

They had OCR iirc but not multimodal model

[vapid slate]

I mean your first Deepseek vision day of release prediction message.

[sharp vortex]

They had leaked their repo tho, so I can cope 🗣️

[jovial kelp]

The quirk of model who aren't being unified with their vision

Deepseek still aren't multimodel and it depend on their seperate OCR model, i guess the training they done to allow better integration between these two seperate models make it hallucinate more on the vision understanding

[thin bramble]

no model is unified with their vision

https://cdn.discordapp.com/attachments/1335677351148916826/1402062060610912410/togif.gif

deepseek simply took it step further and separated the models.

[jovial kelp]

Unified means brought together into a single, cohesive, or functioning whole, characterized by joint action rather than division. It refers to entities, systems, or groups that have been consolidated to act as one, such as a unified team, theory, or structure.

When two seperate architecture/node/block/model being put into one it mean it become unified

[thin bramble]

yeah but that wouldn't be native

so it is not native either way

[rare gale]

Benchmarks on long context performance for pro or flash yet?

[cloud flame]

https://contextarena.ai/

[rare gale]

Hmm, longer scroll than I would like

Probably best to cap at 200k or so tokens then

GPT 5.5 crushing it though

[cloud flame]

It's quite hard test suite, 8needle while previous industry standart was 2needle

[rare gale]

I have been enjoying using this model in Pi though, not bad at all

[cloud flame]

30-40% should be enough @128k to show quality of long context

Wait

Do you also see Claude Opus 4.7 at the bottom of the table?

Huh

[rare gale]

Yes, Opus 4.7 was a big regression in these classic long context benchmarks, they even called it out in the model card

Which is funny after 4.6 crushed them

[cloud flame]

But not like that - I mean what.

[acoustic dune]

Narcissism mixed with flirt

[bright pilot]

Deepseek v4 tomorrow

[raven canyon]

deepseek v5 next week

[sharp vortex]

Why are we in deepseek tmrw era again 💔

[lime moth]

Sorry bros, i must to say something.
In RP
GLM 5.1 >>>>>> deepseek 4 pro.

[peak swallow]

I fully disagree. even glm 5 was better then 5.1 for rp

[opaque coyote]

whoa this is cool - very useful

[hot swan]

damn that's some pretty fucking good scores even for flash
I mean if you remove the various reasoning levels the ranking is just
gpt 5.5
sonnet 4.6
opus 4.6
gpt 5.4
gemini 3.1
deepseek v4 pro
deepseek v4 flash
everyone else

[eternal thorn]

why does it sometimes show the reply to a prompt being done processing but there isn't any text only the reasoning

[woeful jay]

xhigh or high chat ?

[rare gale]

I've been really enjoying this model for agentic coding; maybe it's just because Deepseek's servers are actually relaiable this time round but it's the first time I've thouroughly enjoyed using an OS model for code.

[woeful jay]

when using the api im getting a insufficient balance error

and in the chatroom im getting error 408

ts happening to anyone else? ofifcial deepseek provider only

[oak salmon]

I'm seeing the same issue

[woeful jay]

@deft crow

ok doing byok works but i didnt really want to

[oak salmon]

Yeah, this happened last time when openrouter ran out of credits on their side

[cloud flame]

To be honest, to make 1.6T model bad requires more skill than making it good

[woeful jay]

@deft crow im ngl this is a really bad look for or

[deft crow]

no auto top up from deepseek makes toven a sad boy

already fixed when you first pinged btw

[broken lintel]

it is kind of goofy they don't have auto top up actually

[cloud flame]

I think you need like some kind of web scrapper/LLM to check DS balance every 30 minutes or something like that

[broken lintel]

they have an endpoint to check user balance

https://api-docs.deepseek.com/api/get-user-balance

[cloud flame]

Oh. Even better. But auto agentic top up is another thing lol

[woeful jay]

u da real mvp im sorry king

[woeful jay]

yeah thats understandable

toven tops up with 500k credit limit credit card 👍

[woeful jay]

gump

[woeful jay]

yo this models goated

pro for the price is sooooooooo goated

[tulip estuary]

script or just a ping

[raven canyon]

Has anyone tried DeepSeek fill in the middle completion? https://api-docs.deepseek.com/api/create-completion

[frank wind]

Yeah

I use it exclusively

Uhhh not for coding

[rare gale]

It really is very, very solid

It's very consistent, and capable enough

I think we finally have a good enough model for like best-of-n coding or mass fan out code review agents (or any other agentic workflow equivalents).

The cache hit ratio is pretty incredible too.

[untold locust]

Any way to disable reasoning in deepseek v4 pro ?

I found openrouter params are broken.

[vapid karma]

I keep going back and forth on it. It feels like a smart model that does weirdly dumb things

They did say it was a preview tbf

[frank wind]

Autocompleting is way better than chat

If you know how to write and edit

[frank wind]

Pro is most likely undertrained

Hopefully the Ascends will help them mitigate this

[gusty sphinx]

but what are they filling their middle with? 🤔

ohhh it's just completions but you can optionally provide the suffix prompt as well

that sounds fun actually. i'm tired of talking to these goddamn clankers

[gusty sphinx]

sadly this excludes most users

[untold locust]

is there any param that disable reasoning ?

I see weird pattern it does reasoning sometime and sometime it doesn't

[opaque coyote]

I haven't been using the DeepSeek provider for benching cause of the data training policy. A bit annoying - would save a stack of cash otherwise.

[opaque coyote]

did you try:
extra_body={"reasoning": {"enabled": False}}

[untold locust]

Yes

Problem is it does reasoning sometimes and sometimes it doesn't/.

[opaque coyote]

Hmm weird... maybe a model level issue?

[untold locust]

Yeah , initially i suspected openrouter param mismatch.

But model mention non-thinking mode in hf description.

[jovial kelp]

This model is pretty interesting

Work pretty well to in zed

[peak swallow]

might be a provider issue. its always worked for me with deepseek official provider

[untold locust]

i used asme.

:deepseek

[jovial kelp]

This is interesting

In the past deepseek always has problem with qouta, now with their stack of ascend asicc they seems to have better qouta distribution

Their speed also quite stable at 20-30TPS

I remember last year i need to always use third-party provider because deepseek always got overloaded, making it to slow for me, specially the latency

[bright pilot]

deepseek v4 tomorrow 🙏

[rich ferry]

deepseek v4 yesterday

[woeful jay]

yeah theyre stable at their speeds

which is great

if only they would remove the prompt training

[rich ferry]

I will load 400 million dollars of credits to deepseek if they drop the prompt training

[bright pilot]

its possible btw

just email [email protected]

[rich ferry]

Is that mentioned somewhere in the docs or is this a word of mouth kinda thing

Because that would make DeepSeek pretty much a no-brainer

@bright pilot

[rustic island]

Interesting, skinmed over the docs and that's indeed there

https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html

Users can query basic service information, opt out of data usage for model training, delete their historical data, and more. If you have any claims, requests, or questions regarding the exercise of these rights, please refer to our [Privacy Policy] or contact us at [[email protected]].

[bright pilot]

I've not tried. Shoot your shot

[rich ferry]

I sent them an email, will report back

[sharp vortex]

Imagine after 75% discount ends, you need to manually opt-in to get 75% discount again

-# assume they stop gathering data

[rich ferry]

Honestly, getting that response wouldn't really surprise me

I'm hoping that's not the case but we'll see

[oak maple]

waaaiiiit whaaat

[bright pilot]

They've clearly said they'll get prices down after they get more hardware

I dont understand the confusion with the discount