DeepSeek-R1 and DeepSeek-R1-Zero

[pearl parcel]

I get some reasoning in the response - but no </think> to separate the reasoning from the answer

[bright portal]

ooh Together

Hmm... still working for me, even when I pin to Together

[pearl parcel]

do you want me to print out exactly what's going over the wire?

[bright portal]

Can you try a prompt like "9.11 and 9.9, which one is larger? Please THINK!"?

[pearl parcel]

ok

[bright portal]

it looks correct in the console -- reasoning should not be None if it produced the think tag

[pearl parcel]

definitely it's thinking - it just doesn't seem to be wrapping it in <think></think>

[bright portal]

we detects the think section and pull it into the reasoning field of the delta

So you shouldn't see the <think> tag from our API

[pearl parcel]

take a look at the sample

[bright portal]

Hmmm looks sus for sure, let me double check

in the mean time, if you do stream: true, does it work?

[pearl parcel]

I don't know how to process a stream in python - do you have an example?

[bright portal]

https://openrouter.ai/docs/api-reference/streaming

[pearl parcel]

I switched back to /chat/completions and it worked this time, from DeepInfra

[bright portal]

without streaming right?

[pearl parcel]

without streaming

it worked again, also from DeepInfra

is there a way to force it to Together so I can verify if it doesn't work on Together?

how can I force it?

[bright portal]

https://openrouter.ai/docs/features/provider-routing#example-specifying-providers-with-fallbacks-disabled

Something like this (in TS):

  body: JSON.stringify({
    'model': 'mistralai/mixtral-8x7b-instruct',
    'messages': [
      {
        'role': 'user',
        'content': 'Hello'
      }
    ],
    'provider': {
      'order': [
        'OpenAI',
        'Together'
      ],
      'allow_fallbacks': false
    }
  }),

We will add some Python docs for this soon cc @cerulean lotus

[pearl parcel]

ok it's working now, for both Together and DeepInfra (intermittently, at least)
I will just check for null or empty reasoning and retry

Thank you for helping me

[bright portal]

ur welcome!

Looking forward to the next Dolphin

[pearl parcel]

I need to get it training on 72b, and also I need to get a RL pipeline set up. (Pretty sure Dolphin-R1-24b would be much better if I used RL to tame its <think> block)

[pearl parcel]

2% of responses have no reasoning - That's not too bad

occasionally there's no "choices" in the response

[pearl parcel]

Reason API error: Response ended prematurely
Reason API error: Response ended prematurely
Reason API error: Response ended prematurely
Reason API error: Response ended prematurely
Processing samples:   3%|█▋                                                    | 1514/47532 [32:58<21:57:27,  1.72s/it]Reason API error: Response ended prematurely
Reason API error: Response ended prematurely

getting a lot of these "Response ended prematurely" and a few of

Reason API error: HTTPSConnectionPool(host='openrouter.ai', port=443): Max retries exceeded with url: /api/v1/chat/completions (Caused by SSLError(SSLError(1, '[SSL: SSLV3_ALERT_BAD_RECORD_MAC] sslv3 alert bad record mac (_ssl.c:2570)')))

[rocky heron]

looking

[bright portal]

What client are you using btw?

is it just this raw request.post?

[pearl parcel]

yes

[bright portal]

How do you read the data within the response?

cc @sleek arch

[crystal fjord]

im still, sometimes, getting empty replies from deepseek api provider for r1. is that just cause theyre refusing to reply? sillytav problems? or api problems?

[icy fulcrum]

I'm using the deepseek-ai/DeepSeek-R1 API and need providers with a throughput of at least 25 t/s. However, provider performance constantly changes, and some providers, like Nebius Studio, use different model names (e.g., "deepseek-ai/DeepSeek-R1-FAST"), making it difficult to ensure I'm always using the fastest available option.

Right now, API provider routing only allows filtering by provider name, which isn't enough for my needs. I want a way to automatically route requests to any R1 provider that meets my throughput requirement, without manually updating the provider list.

A better solution would allow filtering providers by throughput, uptime, max context, max output and pricing instead of just excluding them by name.

[formal nest]

We have this now!

Lots of provider routing options here: https://openrouter.ai/docs/features/provider-routing

most recently, we have sort, which lets you sort by throughput, latency, price, etc.

we also have a max_price param: https://openrouter.ai/docs/api-reference/parameters#max-price

let me know if this begins to fill the requests / features you hope to see.

[icy fulcrum]

That works. One suggestion: emphasize displaying each provider's uptime % in the analysis. Uptime % is a crucial metric, but it isn't currently shown on the model page, nor can we sort by it through API calls. Thank you!

[vocal raven]

isn't the whole point of openrouter is that it routes away from downtime without any work on your part

[icy fulcrum]

I see the value of automatic fallback, but in my case, my app relies on chained API calls, data transformations, and long-running processes. Each fallback adds latency and increases processing time. If I can proactively avoid providers with frequent failures, I can minimize disruptions and ensure a more efficient experience for my users. Having uptime data would allow me to make smarter routing decisions upfront, reducing the need for fallbacks in the first place.

[vocal raven]

isn't the whole point of openrouter is that it routes away from downtime without any work on your part

[rocky heron]

Each fallback adds latency and increases processing time

not true! we keep track of how providers are doing and skip them without adding latency if they're unreliable. like @vocal raven said above

[silent gulch]

https://x.com/perplexity_ai/status/1891916644248846789?t=7_5m7rcR2w7GFITF2I2QSA&s=19

Perplexity (@perplexity_ai) on X

Download the model weights on our HuggingFace Repo or consider using the model via our Sonar API.

HuggingFace Repo: https://t.co/9HK9mQGKQ1

X · Feb 18, 2025, 6:24 PM

Can we have R1 1776 from sonar api to Openrouter?

[icy fulcrum]

But how is a provider determined to be unreliable? By detecting failed API calls, right? That means failures have to happen first. And unless OpenRouter calls all providers simultaneously, which would be inefficient, the fallback provider is only called after the first one fails, introducing delays. This has happened to me many times, and it's obvious that making one call takes less time than making two or more sequentially. I’m not saying the fallback system is wrong, I think it’s great. But since a provider can only be detected as failing after it fails, the system could be even more efficient if provider uptime were disclosed and sorting by it was possible through the API.

[rigid nova]

Past returns do not guarantee future returns

On your application side, having logic which retries OpenRouter means you get the routing functionality and the "retry" will then be resolved by OR to a working provider. Only need the one set of API settings in your application.

If you are really mission critical, need to build in logic that means if for whatever reason OR goes down (eg domain hijacking), it goes to some other endpoint for a request.

[icy fulcrum]

Past failures don’t guarantee future ones, but probability matters. A provider with lower uptime has a higher chance of failing, and uptime % is a good indicator of that probability. I already have a retry system, but retrying increases latency for my users. I’m just suggesting a way to reduce the chances of needing a retry in the first place, regardless of whether it’s OpenRouter internally handling the retry or me. Thanks for sharing your thoughts!

[rugged nest]

might be cool to be able to route to providers with at least e.g. 50 tokens/s speed, but use default prio, or price prio

so like a custom speed floor

[rigid nova]

Publishing more data would be great, I agree.

[formal nest]

cc @cedar sentinel

[pearl parcel]

~50% of my requests are coming back with no reasoning, now.

it's getting expensive

[bright portal]

Can you log the provider as well? We return it as provider_name

Or are you pinning it to a known set

[pearl parcel]

on my next subset I will (I don't wanna throw away my progress)

[vocal raven]

you should consider batches

[icy fulcrum]

Hi,

I’ve noticed that R1 is currently returning many errors, as shown on the uptime page. In my app, this causes long-chained API calls to fail, leading to financial losses when expensive processes fail before completion.

To prevent this, I suggest adding an API endpoint to check the current uptime of each model. This would allow apps like mine to monitor uptime and take preventive action. For example, if uptime drops below 90%, I would display a message to users informing them that workflows are temporarily unavailable, preventing them from starting processes that are likely to fail.

This feature would help developers reduce failed requests and unnecessary costs while improving user experience.

Thanks for considering this!

[bright portal]

Are you pinning to any provider in particular?

[icy fulcrum]

My providers array order is: "Fireworks",
"Together",
"Nebius"

Deepseek is ignored.

Fallbacks are enabled. max_price is 8 for both prompt and completion. sort by throughput is enabled.

[rigid nova]

there are ways you can make your application more resilient immediately, staging (send a partial canned response to the user implying thought, or a loadinwheel) while you retry in the background, having a HA fallback mechanism to azure or vertex etc. 98% uptime means 200 of every 10000 will fail, so you know they will fail

using data from OR is slower than having an immediate path for your program to go down in case of API failure

within OR you could set it up to fallback to one of their highly available models, and if it has fallen back as shown in the response, you can request again for the actual R1 response

[icy fulcrum]

Thank you for sharing your ideas!

I have OpenRouter’s fallback mechanism enabled, and I already have a retry system that calls OpenRouter again when an error is detected.

However, the issue is more complex than just handling explicit errors. Some failed API calls don’t return an error at all, making it difficult to detect programmatically. For example, I received the following response:

{ "result": { "id": "gen-1739998011-kBgJfDmW76mQtvtdyK4L", "model": "deepseek/deepseek-r1", "usage": { "total_tokens": 5623, "prompt_tokens": 5619, "completion_tokens": 4 }, "object": "chat.completion", "choices": [ { "index": 0, "message": { "role": "assistant", "content": "[-1]", "refusal": null }, "logprobs": null, "finish_reason": "stop", "native_finish_reason": "stop" } ], "created": 1739998011, "provider": "DeepInfra" } }
The [-1] response is clearly incorrect and not at all what the model should have returned. This isn’t a prompt engineering issue but rather a problem with degraded responses from the model.

These kinds of errors tend to happen more frequently when uptime is lower, suggesting that the providers available are more likely to be operating at degraded capacity rather than failing outright. Having an uptime API would help my app prevent initiating workflows during these degraded states, where failures are harder to detect.

[formal nest]

please vote: #announcements message

[rigid nova]

yeah I get you, you could detect for something that is wrong unless you ever expect answers <4 characters? there is client side logic like that or checking against a local dictionary for if words have actually been returned. character account is simpler

deepinfra hasn't been my favourite but staff wouldn't have brought it back without serious testing. you shouldnt be billed for a response like that imo

[icy fulcrum]

Checking characters might detect some errors, not the ideal solution though.

Yeah I don't have Deepinfra in my providers array, it was called by a fallback. I agree, but what's more important for me is to be able to avoid this kind of errors from occuring. Model uptime tracking via API would help a lot

[cedar sentinel]

we're actually in the middle of revamping our downtime avoidance system, so this is timely!

but this overhaul is more geared towards request statuses returned from upstream (4xx/5xx) and dynamically routing around outages better than we do today

in your case here tho, it looks like our response status tracking won't help much. since it was a "successful" request with 200 status and a non-error finish reason. our uptime stats won't necessarily cover "quality" of outputs

for this sort of thing, we're in the early phases of planning some continuous evals that will penalize providers that consistently output garbage. still more thinking to do on it though, to try and differentiate bad sampling parameters from bad providers

[rigid nova]

my hobby account with sambanova just got API access for the full R1, maybe it is becoming more generally available...

[merry path]

What kind of context limit?

[rigid nova]

Only 4K

[tight jolt]

Supposing that FP8 variants of R1 are a little faster but dumber(?) than those without the "FP8" (so FP16?), is there a way to set the ":nitro" shortcut excluding the FP8 variants?

[rigid nova]

honestly everyone is fp8, this is what deepseek recommends. The others just have not declared it

[tight jolt]

Ah, very interesting. Thanks!
PS: I also found that there's a quantization filter setting: https://openrouter.ai/docs/features/provider-routing#quantization

[rigid nova]

Excellent, they have improved the docs a lot so great that you found it

[earnest wolf]

Fp8 is the highest quant. There is no fp16 or anything

R1 and v3 were natively trained in fp8

[icy fulcrum]

This sounds good. However, please consider adding a GET API endpoint to retrieve:

Overall uptime % of a model – When general uptime is low, the model tends to return more garbage responses from the available providers. This is especially critical for our use case since we rely on long-chained API calls, where failures mid-process result in financial losses and wasted compute. Having this data available via API would allow us to proactively prevent workflows when the model is in a degraded state.

Uptime % of a model for a specific provider – Some providers may perform worse than others, and low provider-specific uptime would allow us to dynamically avoid routing requests to unstable providers. This helps reduce latency caused by sending requests to failing providers and having to retry them multiple times.

This is a very simple feature with huge benefits—uptime data is already available on the frontend, so making it accessible via an API endpoint should be straightforward while significantly improving reliability for developers.

[formal nest]

Thanks for all your feedback, it's very valuable

[inner reef]

it might be you're using json response format, and "[-1]" is a valid json. Using json format for R1 will make the output quality bad, because it's trained to start with <think> tokens.

[vale marten]

Someone opened a PR on the R1 HF repo enabling using prefix=True with a modified chat template https://huggingface.co/deepseek-ai/DeepSeek-R1/discussions/158

cc @peak flame

[dark vigil]

Now if they could just support temperature...

(Via the official Deepseek api I mean)

[pearl parcel]

this is why we have inconsistency.
providers are deviating from the standard template.
neuralmagic's problem.
ideal or no - EVERYONE should use the official template, and workaround that (by inserting a fake <think> token on the client side) if necessary, otherwise chaos ensues.

https://huggingface.co/neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic/discussions/3

my personal workaround is to use the official tokenizer, and tokenize the prompt myself, then call the completions endpoint

[vocal raven]

Surely you mean you apply the chat template
Surely you don't actually convert the text into a series of IDs

[pearl parcel]

reason_tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Llama-70B")

tokenized_input = reason_tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
)
nontokenized_input = reason_tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
prompt_tokens = tokenized_input.numel()
allowed_completion_tokens = max_tokens - prompt_tokens
final_max = min(allowed_completion_tokens, max_tokens)
final_max = max(final_max, 1) - 1
response = requests.post(
    url="http://localhost:4000/v1/completions",
    data=json.dumps({
        "model": "r1-distill-70b", 
        "prompt": nontokenized_input,
        "temperature": 0.5,
        "max_tokens": final_max
    })
).json()

I use tokenized to count the tokens, to properly set the max_tokens
I use non-tokenized to pass to the completions endpoint.

(and it's kind of insane that we need to tokenize the input on the client side like this, just to figure out how to set max_tokens)

[rocky heron]

why do you need to set max_tokens?

[rocky heron]

we're working on doing this automatically

cc @bright portal

[pearl parcel]

because otherwise, a default value (typically 8k) is used.

[rocky heron]

is that too large for you?

[pearl parcel]

I want 32k

[rocky heron]

ah

[pearl parcel]

minus the request tokens

minus 1 lol

final_max = max(final_max, 1) - 1

[rocky heron]

why not just set max_tokens: 32k then? max tokens is the max completion tokens, not the whole thing

[pearl parcel]

because then I get error "you requested blabal but the prompt + 32k is greater than the 32k the model can provide"

[rocky heron]

ohh

ok we should fix this

[pearl parcel]

I worked around it, anyway 🙂

the important thing to note is some providers of R1-Distill-70b are removing the <think> from the prompt template

(which causes it to generate a <think> token - when it wants to)

[rocky heron]

right that is separate and something we're working on

[pearl parcel]

as opposed to the default which forces a <think> token (but then it doesn't generate one)

FYI, I'm getting a number of rambling never-ending <think> (entire 32k with never an answer) I guess it is temperature related

maybe 1% of requests

I can try dropping temperature to .4 or .3

[wheat comet]

H

[vale marten]

Thank you so much for looking into this

[vale marten]

That would be great too

[vale marten]

I talked to the Fireworks people about not being able to prefill with their R1 API. Seems like they will come up with a fix. I mentioned https://huggingface.co/deepseek-ai/DeepSeek-R1/discussions/158
but I am not sure what fix they will come up with.

Mon/Tues

@rocky heron @bright portal @peak flame

[restive wharf]

great job cluing in these amateurs

[pearl parcel]

There's no amateurs here

[clever jolt]

@pearl parcel are any providers using the updated chat template? The ones I tested seemed to still be using the old one

(i.e. the one without <think>)

according to the current tokenizer published by deepseek {"content":"hi","role":"user"} -> '<｜begin▁of▁sentence｜><｜User｜>hi<｜Assistant｜><think>\n'

it doesn't look like R1 providers are doing this yet tho?

[pearl parcel]

Most are using the neuralmagic fp8 out of the box

[clever jolt]

ah ok, just saw your PR on their repo, hopefully they update it

[pearl parcel]

They fixed it

Just a matter of everyone updating their models 😅 and then mitigating the lack of <think> in the response that the UI depends on

[clever jolt]

cool, I'll check again in a few days then

hopefully providers will update their deployments

[vale marten]

For now I am using this template and the completion endpoint (instead of chat completion) in order to prefill thinking... https://huggingface.co/deepseek-ai/DeepSeek-R1/discussions/158

[peak flame]

https://www.perplexity.ai/page/china-launches-underwater-comp-_SRMVJXjQc2LLmJsq2_2zQ

whale literally computing underwater
based?

I don't know where this comes from, but their uptime is starting to be fixed though we still can't purchase credits directly.
https://zzzzzzz.grafana.net/public-dashboards/88296a8e74c14dae8f839c2b9973214b
Few days later edit: Whoops, they're getting intermittent down times again.

[vocal raven]

is it better than https://en.wikipedia.org/wiki/Project_Natick

[formal nest]

FYI, slowly rolling out an update to how we handle thinking generations on ALL R1 models. They will now consistently think, and prefill will now consistenly work.

[wintry dome]

right now, whenever I use Deepinfra or Nebius AI Studio, the reason and context are just two different versions of a reply or sometimes it looks like the first half of the context is in reasoning...but there's no "reasoning" in reasoning.

Deepseek (the source) is still returning it correctly. Does this have something to do with your changes? Did something change that ST hasn't caught up on perhaps?

[rigid nova]

rest in peace hyperbolic:deepseek/deepseek-r1-zero

[vale marten]

Hmm, I'm very interested in knowing what you guys did to make prefill work consistently...

[vale marten]

https://fxtwitter.com/AndrewCurran_/status/1894355918621749402

[sinful crown]

👀 👀 🍿

[tender pawn]

https://i.febryan.me/j9lon.png

Discount on certain time interval.

[junior skiff]

that is interresting source please link

[pale hull]

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

[wintry dome]

So I just got this from DeepSeek. I know this would apply if we user our own API, but I wonder if the discount would apply through OR (assuming using the deepseek provider and not one of the others)

[formal nest]

Yeah if you use your own key through OR, you'd get the discounted pricing.

[wintry dome]

I figured that, but what if we use OR's deepseek provider

[rigid nova]

If you set a Deepseek API key in OR settings, you don't pay anything for the token to openrouter, but you pay 5% of the normal cost as a transaction fee: https://openrouter.ai/docs/use-cases/byok

Make sure you monitor your usage, as when the deepseek provider is unavailable with your key, it will retry with your credits, and you would pay OR for the token.

(setting your deepseek key does not guarantee it will be used, just beware)

[wintry dome]

This doesn't answer my question. Setting aside the BYOK. If I use DeepSeek v3 or DeepSeek-R1 through OR (making sure I use the actual DeepSeek provider when doing so since both models have many providers), will the discount be applied. That's what I'm asking.

[formal nest]

Not at this time. We don't support their pricing discounts

[junior skiff]

because its technically not possible atm or because you just want to make 75% more cash during those few hours ?

if its a technical thing - i can understand that this can get hairy / specially on provider level

[formal nest]

we don't have the code support for it

[junior skiff]

ya that is cool - here is a recommendation that may is a easyer implemenation - weekly refunds of the over charging happening for that 1 provider .. in a given timeframe

so charge full

but refund the delta once a week or a month

should be a fairly easy processing step / and doesnt need to be changed during proxy passing

just fruit for thought

since transaction logs are stored anyway

i think that would be the happy tadeoff

not sure what alex or the rest of the team thinks

[icy fulcrum]

According to this, R1 supports response format. Is this wrong?

[half sapphire]

But we're still getting

their discounts

right?

Especially if I'm using a custom integration key?

[formal nest]

You only get their time-based discount if a generation is going through your own API key, in which case you are billed at what they bill you at, including discounts. And our 5% fee comes from the discounted rate.

otherwise, no, it's the standard non-discounted rates, at all times, through OpenRouter

[half sapphire]

Amazing i'm using my own API key

beautiful, ya'll are goated thank u!!

[vocal raven]

(<t:1740673800:t> - <t:1740616200:t>)

[vocal raven]

i really hope you can implement this soon

[formal nest]

At this time, it is not a high priority item for us, but we are planning to support it at some point.

[opaque veldt]

I can imagine it isn't exactly straight forward, since it is time based, but I wonder if other providers will do similar things as a way to differentiate themselves

[formal nest]

Yeah the thinking on our end is we should have some dynamic pricing capabilities that can normalize across providers / models / implementation details, instead of a one-off

[opaque veldt]

yeah, completely makes sense

as someone mentioned, if it was a big enough deal could do retroactive manual (with scripts) refunds when users would have got the discount

but likely not a huge deal

[vocal raven]

or make a cron job to change the pricing at a fixed time every day

that way the openrouter.ai price is updated too

[opaque veldt]

well that could be prone for issues, since there could be timing issues

[vocal raven]

wdym?

are you saying it might lag behind?

[opaque veldt]

just slight differences between when things actually switch on the provider side, or the cron job failing and the price staying wrong for too long

I personally am not a fan of cron jobs for core infra like that

[vocal raven]

well i don't think anything could fix the provider lagging behind

no way to know that afaik

(response's usage is just token info, no price info)

[opaque veldt]

yeah, though depending on setup, a tighter integration with a provider might be a better option, e.g. where the provider can tell openrouter the price is changed

but that precludes that they care about openrouter 😄

[formal nest]

🎯

[junior skiff]

the retro active refund would be the easy part

as you run it once a month

i dont care if i overpay for 1 provider

if i know i get the money eventually back

going byok defeats a bit of the purpose of holding funds with or

• its just 1 provider anyway / but the savings are nice .. difference if you spend 1000 bucks or 250 on a bigger synth gig

if you get the allocation that is

[pale hull]

DeepInfra added a faster endpoint for R1 at $2/$6 (fp4 quantized) https://x.com/DeepInfra/status/1894866880160244163

DeepInfra (@DeepInfra) on X

🚀 Exciting news! @DeepInfra just dropped Deepseek R1 Turbo—blazing fast at up to 40 tokens per second!

🔥 Runs on Nvidia B200 GPUs
💰 Pricing: $2/$6 per 1M tokens
📍 Hosted in the US 🇺🇸

Try it now on DeepInfra! As always - the best price.

X · Feb 26, 2025, 9:47 PM

[opaque veldt]

max 32k context window though

as chatty in reasoning as r1 is, lol

[tender pawn]

FP4 though, then Nebius is much better.

[formal nest]

are aware of both, will add deepinfra today

nebius wasn’t quite ready for us

[formal nest]

this is up now

[merry path]

Okay now we're talking, 150tps, just need to ge that context window up.

[keen pike]

Now I can't use the original (fp8) deepinfra, they have the same provider id (deepinfra) it seems

It's more expensive and has lower quality

[formal nest]

you can sort by price

https://openrouter.ai/docs/features/provider-routing#provider-sorting

[keen pike]

Any way to just select the provider manually?

[formal nest]

You can select the DeepInfra provider, but unfortunately we don't have a way to let you select the specific endpoint at this time. Your best bet is to sort by price and select DeepInfra, and if you really want to avoid getting any other endpoints, you can disable fallbacks. All of this should be in the docs I linked.

We know this is not a good experience, on our roadmap to fix.

[dry moss]

Can this setting be more agressive?

I have it set to use the cheaper version but the much pricier version is still replying. almost half of the time.

I rather have a generation failing sometimes than paying 10 times the price.

[formal nest]

that's a valid point! the platform defaults tries to guarantee a generation (so it falls back). You should be able to disable fallbacks in your API calls.

[dry moss]

I looked onto it and reached this conclusion:
In my opinion, this is a bit out of touch to the fact a lot of users using open router aren't developers.

The majority of apps don't give this deep level of customization of API calls in their UI. I use mostly for roleplay, but I also saw here, that even extremely widely used apps like cline don't allow to disable fallbacks.

I really think that this should be an option handled on openrouter settings.

I would love if this can be looked into. Because while its true that there is the free version for deepseek, 1- I never take free stuff for granted and 2- It fails more often than the cheapest paid.

[formal nest]

This is good feedback, thank you!

[wintry dome]

Perhaps you should make the "Turbo" variant a sperate model in the list that way it has to be explicitly selected (similar to how Nitro versions are separate though I know that's not the same thing)

[vocal raven]

nitro is no longer separate

everything is in one list

makes sense if you consider that the goal is load balancing

[elfin wharf]

Hi, it looks like sometimes when the reasoning segment in the "raw" R1 response is empty (having things like <think></think>), OpenRouter cannot recognize the text after it as result: it still thinks the tokens after </think> are reasoning tokens.

Also... it could be a good if we could customize the reasoning prefill, like forcing R1 to think in a certain language (prefill <think>OK vs <think>好的).

[limpid wasp]

shame indeed.

[vocal raven]

just gonna drop this for all the non-deepseek providers

$0.55/2.19 is enough for them, why isn't it enough for you

[nimble bobcat]

DeepSeek provider demands collecting prompts, then the pricing is reasonable.
Non-DeepSeek providers with higher pricing, claim they do not collect anything, but who knows?

[vocal raven]

I don't believe the cost here has revenue from data sharing pre subtracted

[clever jolt]

I think the other ones just haven’t got used to the new pricing regime yet. Before it was pricing like o1 and Sonnet, prices where clearly there is a very healthy markup. They probably think “oh well $8 output is cheap then” but it’s not.

[cinder shadow]

Each H800 node delivers an average throughput of ~73.7k tokens/s input (including cache hits) during prefilling or ~14.8k tokens/s output during decoding.
whereas vLLM on a H200 node delivers 1.4k tokens/s output at best at low context. Current inference engines are doing at least 10x lower throughput compared to DeepSeek's inferencing.

[merry path]

With a 500%+ margin at their current prices too, eh

[clever jolt]

Clearly just slapping vllm on a cloud rented instance isn't really good enough to be a competitive service provider.

It just wasn't as noticable before because while I'm sure OAI/Anthropic do have optimized deployments, they are receiving a lot of markup on their $15/output pricing (or $150 for OAI recently lol).

[vale marten]

https://x.com/deedydas/status/1895702508053643560

Deedy (@deedydas) on X

BREAKING DeepSeek just let the world know they make $200M/yr at 500%+ profit margin.

Revenue (/day): $562k
Cost (/day): $87k
Revenue (/yr): ~$205M

This is all while charging $2.19/M tokens on R1, ~25x less than OpenAI o1.

If this was in the US, this would be a >$10B company.

X · Mar 1, 2025, 5:07 AM

[limpid wasp]

this doesn't account for recouping upfront investments and as they stated, most people don't pay (they don't even a paid payment plan on their site, so it's just API calls).

[vocal raven]

upfront investments? this is assuming they lease the h800s

[clever jolt]

I saw some estimates of their value at something like 80b if it was a USA company (though a big range like 1-160b). China business world is extremely different to the USA though.

[junior skiff]

could make if all is billed - webapp and mobile are not billeg

so i think they are breaking even

[solid copper]

paying US tech companies means that at least half of what you paid is going to the pockets of billionaires and shareholders, this wont change anytime soon, THEY aren't willing to change it.

[junior skiff]

was the deepseek integration changed that it wont supply the thinking traces by default anymore ?

[limpid wasp]

was it ever provided "by default"? I always received it with include_reasoning in the reasoning field.

[split cedar]

deepseek's inference efficiency won't be achieved by the other api providers unless they plan to invest a lot of nodes to implement the full EP implementation with separate processing for prefill and decode

[junior skiff]

it was ..

just need to patch st now i guess

[weak salmon]

hey all, been using deepseek-r1 in production through open router.
I have json as the response format and require_parameters = True, and it still routes me to Together which doesnt even support json format. Places like fireworks have a post processing json forcer even for models that dont natively support it so im just wondering why I am still getting routed to non-json-supporting providers or if anyone has gotten around this?

edit: I opened a help thread here https://discordapp.com/channels/1091220969173028894/1346285793953583187

[radiant cape]

I've noticed there is a difference in the response quality between the free and paid versions, is this due to the provider or the money?

[vocal raven]

both

the provider takes no money so you get what you pay for

[radiant cape]

Makes sense

[stark sluice]

which makes sense considering each of their deployment is 320+32 H800s, the standard 8*H200 deployment would be very wasteful as there are a lot of overheads with shifting the weights back and forth

[trail sapphire]

Anyone know provider that serve deepseek r1 zero? hyperbolic has stop serving it.

That model are really good if you trying to get data for whatever you need, yes it's harder to read and understand but for true thinking trace is actually quite unique.

[vale marten]

What were you using it for, if you don't mind me asking

I'm curious

[rocky heron]

Nobody that we know of atm but we’re trying to incentivize one to start up

[formal nest]

👀 https://openrouter.ai/deepseek/deepseek-r1-zero:free

[rocky heron]

^

cc @rigid nova

[peak flame]

I can't get R1 Zero to stop outputting \boxed{}.

[night summit]

does this have tooling ?

[trail sapphire]

It's coming from deepseek original RL, nothing can't be done with it. I think

[trail sapphire]

Thanks 🙏

[trail sapphire]

Some data stuff

[rigid nova]

Even with fiddles of parameters?

[rigid nova]

🥺 🔀🩶

[nimble bobcat]

long reasoning and slow tps; normally it takes 2 mins to finish a turn 🤣

[rigid nova]

new pricing scheme from novita.ai

[woven chasm]

I'm getting an annoying number of 0 tokens out of Nebius for R1. The other providers seem fine.

[trail sapphire]

its more annoying when they deduct the input when we dont get any output

[formal nest]

you are not currently charged for 0 output tokens - #announcements message

will flag to nebius that this is a significant issue

[leaden socket]

Seems like MiniMax is returning 0 output tokens all the time now. Some examples from my logs:

{"status_code": 200, "response": {"id": "gen-1741365750-tmyFfnakL1ErupeZdIGM", "provider": "Minimax", "model": "deepseek/deepseek-r1", "object": "chat.completion", "created": 1741365750, "choices": [{"logprobs": null, "index": 0, "message": {"role": "assistant", "content": "", "refusal": null, "reasoning": null}}], "usage": {"prompt_tokens": 2774, "completion_tokens": 0, "total_tokens": 2774}}}
{"status_code": 200, "response": {"id": "gen-1741365750-IC24jkzeA5Vf625itjxD", "provider": "Minimax", "model": "deepseek/deepseek-r1", "object": "chat.completion", "created": 1741365750, "choices": [{"logprobs": null, "index": 0, "message": {"role": "assistant", "content": "", "refusal": null, "reasoning": null}}], "usage": {"prompt_tokens": 2843, "completion_tokens": 0, "total_tokens": 2843}}}

[peak flame]

Prefill + stop string ["\\boxed{"] to turn R1 Zero into a more conventional non-thinking model (but then what's the point). For RP at least, since the response is easily just (response) \boxed{(response)}. Temp 1.3+. 😆

But I think R1 is still better.

[woven chasm]

Indeed. Deepseek is better now, Can't believe t.

[rocky heron]

why is it better?

[woven chasm]

No 0 tokens. And response time improved

[rocky heron]

this shouldn't be happening - we insure all the zero-output requests for deepseek. did you notice otherwise?

[rocky heron]

from Novita specifically, or from other providers?

[woven chasm]

Deepseeks is not returning 0 tokens, Nebius is.

[formal nest]

@woven chasm can you provide any generation IDs? From our metrics, it's not a large % of requests

Are you setting max_tokens?

[woven chasm]

1024

[formal nest]

Do you see reasoning tokens in your activities page? 1024 is typically too low to get actual completion (non-reasoning) tokens

[woven chasm]

https://openrouter.ai/api/v1/generation?id=gen-1741346738-faG3JKmrg3tW2dTpMsdj

No reasoning tokens.

I Upped the Max_tokens to 32000

[formal nest]

thanks let me look

[woven chasm]

To the same effect.

[formal nest]

can you screenshot your activity tab? It seems in our logs we are getting some tokens back

[woven chasm]

Mmm. As we spoke I'm trying to use it. This is what I'm getting with Deepseek

And this with Nebius

Same prompts. See the difference in output tokens? I believe its thinking, but not passing the results.

[formal nest]

hmm yeah that is strange. If you click to expand on the Nebius generations on the righthand side there, and view the token counts / raw metadata, do you see reasoning tokens?

[woven chasm]

4504 prompt 131 completion,
incl. 128 reasoning

4500 prompt 133 completion,
incl. 130 reasoning

[leaden socket]

Maybe related, I did 500 requests to Nebius, 7 of those have empty content field but full reasoning field.

[woven chasm]

Mmm. I'm seeing error while passing 'max_price', such as this:
{
"max_tokens": 8124,
"temperature": 0.7,
"top_p": 0.95,
"presence_penalty": 0.5,
"frequency_penalty": 1.7,
"stop": ["#"],
"n": 1,
"tools": [],
"tool_choice": "auto",
"max_price": {
"prompt": 1,
"completion": 3
}
}

[formal nest]

what's the error message?

[woven chasm]

Got an unexpected keyword argument 'max_price'

It worked until it didnt

API parameter 'exclude' doesnt seem to work either. This with QWQ paid version

Yup. Reasonin effort as this:
"reasoning": {
"effort": "high",
"exclude": true # Use reasoning but don't include it in the response
}
Doesnt work for me.

AsyncCompletions. create() got an unexpected argument 'reasoning'

The same as with 'max_price'

This parameters work fine:
{
"max_tokens": 8124,
"temperature": 0.7,
"top_p": 0.95,
"presence_penalty": 0.5,
"frequency_penalty": 1.7,
"stop": ["#"],
"n": 1,
"tools": [],
"tool_choice": "auto",
}

[formal nest]

You're using the openai client @woven chasm ? You'll need to pass the reasoning and max_price params etc through extra_body

[woven chasm]

OpenAI client, yes. And this shows my ignorance. Guess I have to find out what extra_body is.

Thank you. Will find out what it is and how to use it. A pointer would be helpful, but you dont have to do that.

[formal nest]

yes one moment

        model="deepseek/deepseek-r1",
        messages=[{"role": "user", "content": prompt}],
        extra_body={
            "include_reasoning": True,
            "max_price": {"prompt": 0.9, "completion": 2.4}
        },
        stream=True
    )```

something like that

so max_price, reasoning, exclude, can all go here

[woven chasm]

Yes, I looked for the information and fixed my json. Today I learned something thanks to you.

So @formal nest, no idea about the '0 tokens' issue I had with Nebius?

[formal nest]

I unfortunately don't see major (>0.5%) occurences of actually 0 completion tokens in my metrics, I can take another look at your account logs though

[woven chasm]

If you could I would be grateful. Its annoying.

[formal nest]

yeah absolutely. Little busy right now but will review later

[woven chasm]

Thanks!

[trail sapphire]

I didn't notice that with deepseek but maybe when I use minimax model few days ago.

[nimble kelp]

request model

[half sapphire]

strictly for R1 and role play purposes

anyone else find the model very agressive

[strange comet]

very aggressive, argumentative, angsty, using the same phrases like "the beer tastes like regret and disappointment", eventually making characters act like lunatics

[upper tapir]

I agree, I choose 12B models over R1 unless I want complete physco

[neat shuttle]

12B?
Are you serious?
For heaven's sake, de gustibus non est disputandum, but only if you use WizardLM-2 8x22B do you realize that you are in a much more advanced era.

[upper tapir]

I retired WizardLM-2 8x22B user, too much positivity

[strange comet]

Minimax has worked pretty well for me. If R1 starts going into a loony place, I rectify it with Minimax for one or two generations and it kind of goes back to normal. I haven't used any other models recently. They're so predictable and almost one-directional.

[rigid nova]

why minimax im curious? i assume for the context length but thought google would be a better option
what have you found its strengths to be

[strange comet]

Gemini is moderated while minimax is not. Otherwise it would be my no 1 model. Minimax passes the strawberry test and follows the prompt and it's slow in progression of the story.

[timid crane]

useful to know - been trying to find the right model to pair with R1.
R1 starts with genius and then progresses slowly towards insanity on so many genres

[peak flame]

R1 Zero Chutes returning 0 output on all prompts now, no error.

[formal nest]

checking

[dry moss]

R1 zero is also a good pair to r1 its way less evil.

the issue is that chutes half of the time don't return reply and when it returns, the reply is in weird formatting.

I don't like wizard 8x22B too, it is smart, but its too... nice.

thanks for the minimax sugestion, i've been using deepseek v3 as a swipe mixup, but the tendency that v3 has to repeat paragraphs do gets annoying.

Edit:

Ok, yeah, no. after 2 weeks trying Both Minimax and Hermes 3 405b, deepseek v3 still delivers the best responses, being the best pair to R1. they both can be as repetitive as v3 if you allow them. But v3 has a greater character personality adherence and understanding of the chat history.

[peak flame]

OK, responses are back.

[formal nest]

thanks for flagging

[strange comet]

Can you even use Zero for RP? I get nonsense each time. Not even proper sentences. Also try Hermes 3 405b to mix things up. Even better than Minimax, depending on what you want.

[peak flame]

Zero turns into a normal model with a prefill to skip the thinking block and ["\\boxed{"] stopping string. Word salad is rare to me (once in 25 today), goes away on swipe. Didn't try long context. TC without instruct sequences happens to be okay too. "Issue" is there are better RP models.

[half sapphire]

Is there a way to prompt or prefill to somewhat reliably prevent it from thinking too much or at all

[dry moss]

openrouter chat has a slider for thinking slider token budget

if the app your're using don't support this, I think the easiest way to avoid it overthinking is making the prompt as clear as possible so it doesn't have many "but wait!" moments.

I'll assume chub, (mutual servers) So make your system prompt and anything that sets (How it should behave) as simple and without contradictions as possible. You can be descriptive as you want in character details and appearance, but in telling it HOW it should behave, be as simple as possible.

This is not only to avoid r1 overthinking, and hitting chub's 2048 output limit while still on the thinking phase, but because R1 focus A LOT on HOW it should behave, put too much "how" and it will overshadow the rest of your character. (Afterall thats exacly what the thinking phase does, focus on "how" to respond)

@half sapphire (forgot to quote reply)

[half sapphire]

Wait how is that applied on the API side

like is that a param?

[dry moss]

that, I have no idea, I know there is this feature on openrouter chat when selecting r1, so the model probably supports it.

But in any case, I think jailbreaks only make it spend more tokens reasoning, and prefill sometimes just don't work or make it skip the reasoning phase completely.

[peak flame]

👀 DS docs says "soon"

[half sapphire]

👀

[solid copper]

question, the endpoints that don't disclose if they use fp4 and fp8 is because you may be routed to a gpu that can use either quantization? Or simply because they don't provide this information to openrouter?

[formal nest]

latter, we don't know

[cinder shadow]

Fireworks has dropped their pricing to $0.55/$2.19 on their new basic deployment (deepseek-r1-basic) https://fireworks.ai/blog/fireworks-ai-developer-cloud
old $3/$8 pricing applies to their existing fast model (deepseek-r1)
cc @formal nest

[rugged nest]

oh, very nice!!

[clever jolt]

thats 2 providers now with a "fast" end point that it more expensive. Any idea what the difference is?

[earnest wolf]

faster token generation speed

[clever jolt]

yeah I know that I was more thinking whats the difference in underlying deployment

[vocal raven]

GPU count, or quantizing then marking up just for stability

[clever jolt]

yeah thats the thing though, more gpu should mean faster yes, but also larger scale deploy so also cheaper/more efficient

[weak salmon]

hey anyone else not getting any content from r1 calls? Using fireworks. Reasoning works, then i get the "stop" finish reason, then absolutely no content. this is no matter what i do with prompting

Just confirmed, the content is being sent in the reasoning field from openrouter

[bright portal]

Which provider is this btw? -- sometime, reasoning might eat up all the context

[rocky heron]

Works for me:

Fireworks

[weak salmon]

okay let me keep investigating here and make sure my ducks are in a row

json mode btw

I wasn't changing anything just fired up the ol app and getting this prob. Some more details,
heres my object im sending

  response = client.chat.completions.create(
      model="deepseek/deepseek-r1",
      messages=...,
      stream=True,
      response_format={'type': 'json_object'},
      extra_body={
          "provider": {
              "require_parameters": True,
              'order': [
                  'Fireworks'
              ],
              'allow_fallbacks': False
          }}

then what happens is, i start recieving reasoning data that is just my json object that should be in content. I dont recieve any actual reasoning. Then i receive no content, and a finish reason "stop". This is consistently happening. am i crazy? lol is it replicable?

I commented out json and switched to SambaNova and it resolved.
I switched to Fireworks and disabled json format and it also worked.
It appears to be an issue with Fireworks's JSON mode (they are the only people i think that have json mode for deepseek through some proprietary method)
to be clear the issue is: 1- no actual reasoning data and 2 content appearing in reasoning, when in json mode

[jovial flame]

@weak salmon I believe thats an error since json_format wont really work with R1. We should disable allowing users to send that.

[weak salmon]

@jovial flame That would be extremely disappointing, considering fireworks literally supports it

[jovial flame]

Haha you are right

Will look into it first thing tomorrow and get it fixed

[pale hull]

It seems like there is currently no way to specify routing to the non-basic Fireworks?
This is also the same for other multi-version cases, so I think the "provider label" could be exposed to the endpoints API, and specifiable as the routing parameter.

[weak salmon]

haha youre the best i really appreciate it. It seems they came out with a new R1 model so it may have mixed their stuff up.

[frozen dawn]

I enconter the same problem, any one have any idea how to use fireworks fast version of deepseek R1 model?

[vocal raven]

still waiting for offhours discounts

[rigid nova]

The crypto providers are still setting themselves up to float, but there are now very cheap options like fireworks who have a basic endpoint now

[vocal raven]

yeah tbf the cheap ones are getting faster and the fast ones are getting cheaper

[rigid nova]

Klusterai and others do batch processing for a discount

[plucky eagle]

Hello everyone, my question may seem simple and obvious, but I'm just getting the hang of it, so far I can't figure out what's the matter. For some reason, R1 is constantly randomizing the answers in terms of reasoning. He can process the request correctly and separately issue "reasoning", separately "content". Maybe "reasoning" just doesn't fill it out and write a response normally. Or maybe, for some reason, insert the reasoning itself into the "content", i.e. literally the course of your thoughts, has anyone encountered this? Calling like this:

first_model = "deepseek/deepseek-r1"
models_to_try = ["openai/o3-mini-high", "anthropic/claude-3.7-sonnet:thinking"]
#
extra_body = {
"models": models_to_try,
"provider": {
"order": ["Fireworks", "Novita", "Nebius", "DeepSeek"],
"ignore": ["DeepInfra"],
"allow_fallbacks": False
}
}

response_data = await global_variables.ai_client.chat.completions.create(
model=first_model,
extra_body=extra_body,
messages=data_for_requests["messages"],
temperature = 0.4,
max_tokens=20000
)

[vocal raven]

don't set model and models at the same time

just make one models sorted by preference

aside from that idk why content and reasoning would be inverted...
maybe you should go to https://openrouter.ai/activity and send a screenshot and request id of one of the requests with inverted content/reasoning

[weak salmon]

hey it works again, just wanted to say thanks for your fast response if you did that! Appreciate it

[plucky eagle]

it seems that I just accidentally found the reason, the fact was that the request was not in English, as soon as I started writing exclusively all requests in English, everything fell into place.😐

[weak salmon]

edit - sorry for the early props , i was wrong, its still broken in the same way ~ Im going to open up a thread. my whole app is just kinda down because of this

[jovial flame]

yes somed way to repro would be good

[weak salmon]

okay just opened a thread and sent repro object from my tests. Going to hook up to fireworks directly and see if its a prob with them on their side too

[timid crane]

With Targon (from ST) I am getting broken responses - returning only
<Tool Response>
Chutes is working perfectly.
First time I have seen this issue from Targon. Others I have spoken with seeing the same.

[lone sky]

V2 wen?

[vocal raven]

Jul 17, 2024

[lone sky]

i think i only get 1% of the humor. why such a specific date? I don't get ittt 😭 🥜🤏🧠

[vocal raven]

that was when deepseek v2 launched

sorry

im being pedantic

[lone sky]

I read May 2024 on googe tho :(

[vocal raven]

about the difference between r2 and v2

or v2 and v4

[lone sky]

I'm still behind u i think

[vocal raven]

hm idk

i got the date from hf commit dates

[lone sky]

Is deepseek zero more knowledgeable than r1?
I just gave it a knowledge test for eye colors of characters, and it consistently did better than R1.
What else is deepseek zero better at? wow.

[sinful crown]

Being less censored

[lone sky]

Hmm. Maybe that's it. I was wondering about more niche things but I suppose thats a plus.

It has the highest unncensored score on dubesor.de.. I wonder what the Venice chatbot would get

[trail sapphire]

It's build upon base without instructions, so it's the purest form of RL

When you didn't put human bias the model optimize it self, better in many aspect but worse in aspect that human care.

Because it's hard to read and hard to understand but it's actually high quality

To bad hyperbolic stop hosting it, and I don't like chutes hosting

[proven atlas]

If I remember correctly, DeepSeek Zero is a base model instead of a instruction-tuned model. So it should be better at completing sentences.

But last time I tried, it just starts solving math equations randomly out of no where.

[amber stirrup]

Zero is almost further back than a base model

It's a base model run through totally unstructured reinforcement learning on verifiable problems.

I think they very gently showed it what a thinking step looks like, but that's about it. It will apparently switch languages mid-thought sometimes, or use tokens seemingly unrelated to words

[proven atlas]

normal R1 will also switch languages mid-thought, but it was mostly mitigated via penalty in reward during post-training

[amber stirrup]

Interesting, haven't seen it do that yet. Interpretability is cool, but I also love the idea of seeing raw thoughts

[limpid wasp]

I didn't notice that. It does use messier thought chains for sure (not rephrased to be more human friendly), but it was completely usable in my testing. #1330820209812050002 message It understood and tackled all problems, without producing gibberish.

[amber stirrup]

They mentioned it in their paper, but I'm not sure how common it is

[frail oxide]

Someone leak when R2 releases

[vale marten]

soon I hope

there might be some leak on twitter by the usual accounts

[jaunty drum]

🤔

[weary crown]

https://www.bloomberg.com/news/articles/2025-05-28/deepseek-unveils-update-to-r1-model-as-ai-race-heats-up

(paywalled)

[earnest wolf]

"As ai race heats up" I dunno man

It's been pretty hot for a while

It's not heating up

[queen basalt]

Is new R1 available on DeepSeek API?

[sinful crown]

Finally, DeepSeek R1 2!

[earnest wolf]

Yes

[queen basalt]

Need on OpenRouter

[earnest wolf]

@formal nest deepseek did a silent upgrade (again)

[formal nest]

[earnest wolf]

Source: here

The announcement says that the new R1 is live

[earnest wolf]

Are you crying or are you laughing? /tease

[formal nest]

yeah we hit their deepseek-reasoner endpoint so you can just route to deepseek direct and use the upgrade

[sinful crown]

Did DeepSeek make an endpoint for the old one?

[earnest wolf]

Nope

[sinful crown]

Oh, no

[earnest wolf]

You don't like the new one?

[sinful crown]

I have an unhinged roleplay bot running on R1 for a server, just scared that this will change the personality

[earnest wolf]

As an alternative, you can use the free DeepSeek R1 providers

Or cough up more money and use the other paid providers

Test if the personality changed before switching providers

[prisma goblet]

did they also update the api or just their web platform?

[earnest wolf]

Both

[sinful crown]

The inner reasoning for this new R1 is very interesting to watch

It seems more articulate than before

[candid ice]

they improved the CoT alot

[rugged nest]

can't wait for the new thread here 🤓

[queen basalt]

of course it still thinks it's ChatGPT lol

[slim remnant]

Is the new version still open-weights? Presumably we'll get it via 3rd party providers at some point?

[earnest wolf]

Weights aren't out (yet)

[half sapphire]

where does it say it's out for API 👀

[uneven gust]

Hope there's gonna be more than just upgraded R1

It's alright but not 2.5 pro level

[vale marten]

According to everyone, R1.5 feels different. I want more objective data showing performance

[uneven gust]

I haven't tried that same prompt with regular R1

[stray locust]

need bench before & after today

[uneven gust]

Parasail's R1 seems slightly worse

Nvm not slightly

Some chess pieces can't move at all

Definitely an improvement

[celest pilot]

Wonder if the writing capabilities will be better or worse.

[jaunty drum]

https://huggingface.co/deepseek-ai/DeepSeek-R1-0528

it's up, no details yet

[woven quail]

They always update the details later

[half sapphire]

Ok r we sure the API updated on their Provider?

@formal nest sorry for the quick ping, seeing if u know if their API updated to the new one

[formal nest]

i don't have full confirmation really that it's their new model

who knows

[clever jolt]

it would be in-line for them to just update the API endpoint, thats what they did before every time, an in-place update.

big question i guess is did they do it yet

[formal nest]

https://openrouter.ai/deepseek/deepseek-r1-0528

parasail soon

[uneven gust]

Yeeeaaahhh

Does that mean no r2 or v4?

:(

Cuz ye, the model is cool and all but

Anthropic and google have gone way beyond

[tacit vortex]

Yay

Something happened, the nothing ever happens crowd in disbelief right now

[tacit vortex]

I’m just waiting for dubesor lol

[blissful peak]

I think it is same as before

[gaunt rose]

2 providers already up

[earnest wolf]

They did

[rugged nest]

same

[limpid wasp]

well, that'll be a few hours.. I estimate 12ish or so

[earnest wolf]

Can you ping me when you finish the benchmarks? (dont worry if you forget to ping me lol)

[formal nest]

do we need an @ dubesor-benchmark-enjoyer role

you can only get the role if you give dubesor an OR key to bench with

[tacit vortex]

Yes we do

[tacit vortex]

That’s why you’re the GOAT! The GOAT!

[limpid wasp]

🤣 Naw I have rejected keys in the past. But if you wanna contribute data to e.g. the chess leaderboard, just play using your own keys... I incorperate the data of any finished match and thus far I have spent ~$300 on the chess data alone 😦

[blissful peak]

300 $ is insane

[limpid wasp]

there's ya ping

[earnest wolf]

Thank ee very much

[wind tendon]

When will OpenRouter support tool calling on R1-0528 like DeepSeek API does? @rocky heron 🙏🏻

[crystal anvil]

DeepSeek is not able to review images right?

[earnest wolf]

Correct

[gaunt rose]

Is there tool call support with R1? I know R1 v2 supports it, but some folks are saying it isn't supported with OR yet. Is this true?

[earnest wolf]

R1 the model doesn't support tool calling afaik

R1 v2 explicitly mentions tool calling support though

[gaunt rose]

Yup, I'd love to be able to use its tool calling capabilities through OR

[cedar sentinel]

sorry for resurrecting a dead thread, but fyi we have this on a per-endpoint basis now:
https://openrouter.ai/docs/api-reference/list-endpoints-for-a-model#response.body.data.endpoints.uptime_last_30m

https://openrouter.ai/api/v1/models/anthropic/claude-sonnet-4/endpoints

[proven atlas]

I just got this email which looks interesting

[earnest wolf]

Ooo

[tacit vortex]

Nice

Is it the 1bit architecture or is it the Xeon stuff

[proven atlas]

I am getting more info from them. probably will make a blog post / newsletter post on this. will share more when i have info. they have api but need to request.