Inference is impressively fast. But what about quality? In the Kimi vendor verifier (https://github.com/MoonshotAI/K2-Vendor-Verifier/), Together has one of the highest tool call failure rates (>300 failures over the benchmark, compared to 0-2 for the official API, groq, SiliconFlow, and Infinigence).
I don't know anything about Together quality in general, but the specific technique discussed here (speculative decoding) has no impact on the quality of generations. So you should be able to apply it to whichever model you want, and see the advertised speedup while retaining the quality of your base model.
> the specific technique discussed here (speculative decoding) has no impact on the quality of generations
I don't see why that would be true. As I understand, the verifier is checking if the tokens are good-enough, not if they're the exact same tokens it would have selected. The predicted tokens could be consistently slightly worse, which could have a cascading effect to make the overall output a lot worse.
I didn't know this! I've always thought speculative decoding was "if p(draft_token) > threshold, use it". You made me go read how it actually works and it's pretty neat!
That said, I still think some providers are cheating. Please correct me if the test below is flawed.
I generated texts at temperature = 0 vs temperature = 2. At high temperature, the distributions effectively become flatter, meaning the difference between real and draft effective distributions (the D_LK used in theorem 3.5 of 2211.17192) becomes smaller. When T=2, the model speaks complete gibberish, so the effective distribution must be pretty flat. This should mean fewer rejections --> a lot faster speculative decoding. Yet, I see no increase in throughput at all...
If you compare "schema validation error count" plus "Count of Finish Reason others" then SiliconFlow and Infinigence is in the same bucket too. Maybe their API layer detected incorrect tool call and set finish reason to something else?
IMO this likely is what you get from running the model correctly as-is (i.e. using the same weight and activation dtype), so Together is not bad.
Moonshot AI themselves and Groq likely uses some sampler tricks to eliminate schema validation errors.
So really the only thing this shows is: Nebius, Chutes, AtlasCloud could be running something else (for example further quantized model). Or bugs.
Fair point. If Moonshot is holding back the true weights or inference techniques that affect correctness, then providers including Together should call them out on that. I for one would stop using Kimi if that is the case.
Anyway, Novita is doing significantly better on the vendor verifier chart than Together, so the low quality must be partially Together's fault at least.
>a faster speculator (also known as the draft model) proposes multiple tokens ahead, and the target model verifies them in parallel in a single forward pass
TIL. Bit of an aha moment - never understood till now how a big model can verify faster than it can generate
As with almost everything else in CS, it's a tradeoff. Pre-fill is compute bound, decoding is memory bandwidth bound. Speculative decoding works when the draft model is more often right that wrong, because most architectures have a lot more compute, compared to memory bandwidth.
At first glance, this reminds me of how branch prediction is utilized in CPUs to speedup execution. As I understand it, this development is like a form of soft branch prediction over language trajectories: a small model predicts what the main model will do, takes few steps ahead and then verifies the results (and this can be done in parallel). If it checks out, you just jump forward, it not you take miss but its rare. I find it funny how small-big ideas like this come up in different context again and again in history of our technological development. Of course ideas as always are cheap. The hard part is how to actually use them and cash in on them.
A lot of optimizations in LLMs now are low hanging fruits inspired by techniques in classical computer science. Another one that comes to mind is paged KV caching which is based on memory paging.
Will need some time to go through the details, but it’s increasingly rare to see teams consistently delivering meaningful improvements in the open. Impressive work!
> Built on top of Together Turbo Speculator, ATLAS reaches up to 500 TPS on DeepSeek-V3.1 and up to 460 TPS on Kimi-K2 in a fully adapted scenario — 2.65x faster than standard decoding, outperforming even specialized hardware like Groq
You'll see Groq averaging 1,086tps vs Together doing 59tps. Groq and Cerebras often feel like the only games in town. I'd love that to be different (because I'd like more models!), but nobody else is coming close right now.
Comparing how quickly gpt-oss-120b runs gives a broader picture: https://openrouter.ai/openai/gpt-oss-120b -- Vertex (Google) and SambaNova do pretty good on it too, but still, the difference between a top provider and an also-ran is giant.
> I'd love that to be different (because I'd like more models!), but nobody else is coming close right now.
I'm currently on the Cerebras Code subscription for like 50 USD a month because it more or less makes the rate limits I used to deal with other platforms disappear (without making me spend upwards of 100 USD paying per token): https://www.cerebras.ai/blog/introducing-cerebras-code
At the same time, their Qwen Coder 480B model is fine but I still find myself going for Claude or GPT-5 or Gemini 2.5 Pro for more complex issues (or ones where I need good usage of Latvian language), at least for programming tasks it'd eventually be super cool if they could offer more models.
Or have some sort of a partnership with Anthropic or whoever, because getting my questions answered at around 500-1500 TPS is really, really pleasant, especially for agentic use cases with code modifications, even if I still bump into the 128k context limits occasionally.
2x jump overnight. new LPU hardware? I checked the speed for groq's gpt-oss-120B, Llama4-maverick, and Llama4-scout; none of them had a noticeable change this month
There's another angle to this comparison. Groq and Cerebras use custom chips, but I'm not sure about Together. In this case, Together is sharing results based on the B200 GPU. Another important point is the accuracy of these speed-ups compared to the baseline model. It's known that such tricks reduce accuracy, but by how much? Kimi has already benchmarked several providers. https://x.com/Kimi_Moonshot/status/1976926483319763130
Not just custom chips, but custom chips which derive much of their performance from enormous amounts of SRAM. There's no denying that approach is fast, but it's also incredibly expensive, and SRAM scaling has slowed to a crawl so it won't get much cheaper any time soon.
Wonder if it’s prompt caching? OpenRouter is (I guess) just reporting actual throughput, where presumably groq is reporting a from-scratch figure? Just a guess tho.
But Groq/Cerebras are hardware accelerators. It's an unrelated optimization. I wouldn't be surprised if they could also use speculators (today or in the future).
Inference is impressively fast. But what about quality? In the Kimi vendor verifier (https://github.com/MoonshotAI/K2-Vendor-Verifier/), Together has one of the highest tool call failure rates (>300 failures over the benchmark, compared to 0-2 for the official API, groq, SiliconFlow, and Infinigence).
I don't know anything about Together quality in general, but the specific technique discussed here (speculative decoding) has no impact on the quality of generations. So you should be able to apply it to whichever model you want, and see the advertised speedup while retaining the quality of your base model.
> the specific technique discussed here (speculative decoding) has no impact on the quality of generations
I don't see why that would be true. As I understand, the verifier is checking if the tokens are good-enough, not if they're the exact same tokens it would have selected. The predicted tokens could be consistently slightly worse, which could have a cascading effect to make the overall output a lot worse.
4 replies →
I didn't know this! I've always thought speculative decoding was "if p(draft_token) > threshold, use it". You made me go read how it actually works and it's pretty neat!
That said, I still think some providers are cheating. Please correct me if the test below is flawed.
I generated texts at temperature = 0 vs temperature = 2. At high temperature, the distributions effectively become flatter, meaning the difference between real and draft effective distributions (the D_LK used in theorem 3.5 of 2211.17192) becomes smaller. When T=2, the model speaks complete gibberish, so the effective distribution must be pretty flat. This should mean fewer rejections --> a lot faster speculative decoding. Yet, I see no increase in throughput at all...
3 replies →
If you compare "schema validation error count" plus "Count of Finish Reason others" then SiliconFlow and Infinigence is in the same bucket too. Maybe their API layer detected incorrect tool call and set finish reason to something else?
IMO this likely is what you get from running the model correctly as-is (i.e. using the same weight and activation dtype), so Together is not bad.
Moonshot AI themselves and Groq likely uses some sampler tricks to eliminate schema validation errors.
So really the only thing this shows is: Nebius, Chutes, AtlasCloud could be running something else (for example further quantized model). Or bugs.
Fair point. If Moonshot is holding back the true weights or inference techniques that affect correctness, then providers including Together should call them out on that. I for one would stop using Kimi if that is the case.
Anyway, Novita is doing significantly better on the vendor verifier chart than Together, so the low quality must be partially Together's fault at least.
1 reply →
>a faster speculator (also known as the draft model) proposes multiple tokens ahead, and the target model verifies them in parallel in a single forward pass
TIL. Bit of an aha moment - never understood till now how a big model can verify faster than it can generate
As with almost everything else in CS, it's a tradeoff. Pre-fill is compute bound, decoding is memory bandwidth bound. Speculative decoding works when the draft model is more often right that wrong, because most architectures have a lot more compute, compared to memory bandwidth.
So with a 4x speed-up, Together will give us at least 2x lower price for top-end models, right? :)
Do I understand this right?
A light-weight speculative model adapts to usage, keeping the acceptance rate for the static heavy-weight model within acceptable bounds.
Do they adapt with LoRAs?
At first glance, this reminds me of how branch prediction is utilized in CPUs to speedup execution. As I understand it, this development is like a form of soft branch prediction over language trajectories: a small model predicts what the main model will do, takes few steps ahead and then verifies the results (and this can be done in parallel). If it checks out, you just jump forward, it not you take miss but its rare. I find it funny how small-big ideas like this come up in different context again and again in history of our technological development. Of course ideas as always are cheap. The hard part is how to actually use them and cash in on them.
A lot of optimizations in LLMs now are low hanging fruits inspired by techniques in classical computer science. Another one that comes to mind is paged KV caching which is based on memory paging.
No barrier to entry whatsoever? Backprop on the speculative decoding weights during inference to improve their accuracy on a per application basis?
Cool hack though, kudos. Wonder if they can make Groq or Cerebras do the same thing?
Will need some time to go through the details, but it’s increasingly rare to see teams consistently delivering meaningful improvements in the open. Impressive work!
I can't help but wonder how much longer we'll see this work shared openly.
> Built on top of Together Turbo Speculator, ATLAS reaches up to 500 TPS on DeepSeek-V3.1 and up to 460 TPS on Kimi-K2 in a fully adapted scenario — 2.65x faster than standard decoding, outperforming even specialized hardware like Groq
and yet, if you click on: https://openrouter.ai/moonshotai/kimi-k2-0905
You'll see Groq averaging 1,086tps vs Together doing 59tps. Groq and Cerebras often feel like the only games in town. I'd love that to be different (because I'd like more models!), but nobody else is coming close right now.
Comparing how quickly gpt-oss-120b runs gives a broader picture: https://openrouter.ai/openai/gpt-oss-120b -- Vertex (Google) and SambaNova do pretty good on it too, but still, the difference between a top provider and an also-ran is giant.
God I love OpenRouter.
> I'd love that to be different (because I'd like more models!), but nobody else is coming close right now.
I'm currently on the Cerebras Code subscription for like 50 USD a month because it more or less makes the rate limits I used to deal with other platforms disappear (without making me spend upwards of 100 USD paying per token): https://www.cerebras.ai/blog/introducing-cerebras-code
At the same time, their Qwen Coder 480B model is fine but I still find myself going for Claude or GPT-5 or Gemini 2.5 Pro for more complex issues (or ones where I need good usage of Latvian language), at least for programming tasks it'd eventually be super cool if they could offer more models.
Or have some sort of a partnership with Anthropic or whoever, because getting my questions answered at around 500-1500 TPS is really, really pleasant, especially for agentic use cases with code modifications, even if I still bump into the 128k context limits occasionally.
Interesting, if you take a look at the median throughput chart [0], groq goes insane after 7th Oct. Wonder what happened.
[0] https://openrouter.ai/moonshotai/kimi-k2-0905/performance
2x jump overnight. new LPU hardware? I checked the speed for groq's gpt-oss-120B, Llama4-maverick, and Llama4-scout; none of them had a noticeable change this month
Heavy quantization
1 reply →
There's another angle to this comparison. Groq and Cerebras use custom chips, but I'm not sure about Together. In this case, Together is sharing results based on the B200 GPU. Another important point is the accuracy of these speed-ups compared to the baseline model. It's known that such tricks reduce accuracy, but by how much? Kimi has already benchmarked several providers. https://x.com/Kimi_Moonshot/status/1976926483319763130
> It's known that such tricks reduce accuracy
AFAIU, speculative decoding (and this fancier version of spec. decoding) does not reduce accuracy.
3 replies →
> Groq and Cerebras use custom chips
Not just custom chips, but custom chips which derive much of their performance from enormous amounts of SRAM. There's no denying that approach is fast, but it's also incredibly expensive, and SRAM scaling has slowed to a crawl so it won't get much cheaper any time soon.
1 reply →
> You'll see Groq averaging 1,086tps
What I don't understand is, Groq reporting 200tps for the same model: https://console.groq.com/docs/model/moonshotai/kimi-k2-instr...
OpenRouter numbers look fishy.
Wonder if it’s prompt caching? OpenRouter is (I guess) just reporting actual throughput, where presumably groq is reporting a from-scratch figure? Just a guess tho.
groq is quantizing, even though it's not labeled as such on openrouter (super frustrating)
Do you have a source for that? They are pretty close to the ref implementation on moonshot’s ranking
1 reply →
But Groq/Cerebras are hardware accelerators. It's an unrelated optimization. I wouldn't be surprised if they could also use speculators (today or in the future).
>Groq and Cerebras often feel like the only games in town.
SambaNova should be similar...they've got a similar specialized hardware approach
Do these numbers compare performance at the same cost?
You can see the cost in the links, and the answer is “pretty much” for the consumer. The backend maths, no idea.
Great, my slop memes can come out much faster now. This is the future of the world economy!