I think your demo need more realistic thinking logs because thinking usually burns at least 2x to 3x of tokens of the code and for harder tasks much more.
It's interesting how even 5 tok/s is still much faster than you'd typically type, but feels glacially slow for an agent.
On the other hand, I've been using Mimo and Minimax a lot recently. They routinely reach 100-150 tokens per second and that feels too fast, to the point where it's hard to keep up with what it's actually doing. Great for subagents though.
They routinely reach 100-150 tokens per second and that feels too fast, to the point where it's hard to keep up with what it's actually doing.
There is no way you can follow what is going on even at 30 tokens per second. Maybe you can maintain a rough idea of what is going on for some tens of seconds but that is probably about it. Follow it in any detail, no chance. Reason about what you read, absolutely no chance.
800 tok/s — Cerebras-class, where the bottleneck is your eyeballs
I do not understand why they say this. I am not sure if it is even true, 800 tokens sounds like a page of text and I would assume you can look at one page per second without hitting any limitation of your eyes. Or is the resolution of the human not good enough to see an entire page at once and you have to scan it with the fovea? Scrolling text might of course hit the temporal resolution limit. But why does this even matter, your brain can not process anything close to the amount of information your eyes can take in.
This is awesome!! I use Cursor and I've been trending towards medium thinking models as much as possible - I don't like the dev cadence with something like opus 4.7 (thinking: very high) (great for some tasks, like complex plans). Eventually I'd like to make my way to open models and open harness, and this tool or something like it could help me understand what performance I'd need for productive work - bookmarked!
Cool visualization, but most of the token generation in my sessions doesn't go to output code or even the text I see. Reasoning tokens make up most of the output. That can only occur after processing the input files and context.
For non-trivial work I go through hundreds of thousands of tokens (combined prefill + tg of course) before even getting to some useful text output.
I mostly use LLMs for exploration and studies, rarely code generation. Prefill matters heavily for this. Even in the high hundreds or low thousands prefill rate I spend a lot of time waiting on the LLM (doing other things, not twiddling thumbs)
I'm flashing back to using a 1200 baud modem when the world was on 28.8k. Modems are much more regular-looking, though, since each character is a character. Unless you count color changes and such, which you only really notice at 1200...
Token/sec only makes sense once you tell me three four things:
1. decoding t/s, that is, when the model is generating text in the autoregressive fashion.
2. prefill t/s, that is, prompt processing speed.
3. What is the slope of those two numbers as the context size increases. An implementation that decodes at 50t/s with 2k context but decodes at 7t/s at 100k context is going to be a lot less useful that it seems at a first glance for a big number of real world use cases.
4. What's your use case? Reading a huge text and then having a small output like, fraud probability=12%? Or Reading a small question and generating a lot of text? This changes substantially if a model is usable based on its prefill/decoding speed.
For instance my DS4F inference on the DGX Spark does prefill at 350 t/s and at 200 t/s on already large contexts. But decodes at 13 t/s.
On the Mac Ultra the prefill is like 400 t/s and decoding 35 t/s.
The two systems can perform dramatically differently or almost the same based on the use case. In general for local inference to be acceptable, even if slow, you want at least 100 t/s prefill, at least 10 t/s generation. To be ok-ish from 200 to 400 t/s prefill, 15-25 t/s generation. To be a wonderful experience thousands of t/s prefill, 100 t/s generation.
Agreed. Prefill kills me for local model work. The model reads much faster than it writes, but I'd love to get a sense for how fast the model can read large source conversations.
This reminds me of when I signed up for cerebras to try it out and dumped $20 in and hooked it into opencode and the speed was truly insane. But my one session burnt through like $15 of that in seemingly a matter of minutes. I've since used those really high tok/s options for specific application use cases, but would not advise as a coding agent. Much harder to catch issues when it is moving a million miles an hour and then it is too late and it has already spent a ton of tokens.
It really depends. With the new "thinking" models they usually spend some time before writing the final answer. If they "think" for 1k tokens, that's a minute of spinning wheel you're gonna see for each question. Add that to the prompt processing, and diminishing speeds as context increases, and it becomes really slow for longer sessions.
The nice thing about DeepSeek and off-memory streaming is that you ought to be able to batch multiple sessions of it in parallel. Each individual session would slow down from streaming incrementallly more active weights from disk, but your total tok/s would ultimately only be limited by compute. Other models have trouble doing this, because the KV cache takes too much space in RAM (and increases wear-and-tear if stored on disk) even for somewhat limited context.
30tok/s looks fine when you're just streaming code, but the issue is that there's a lot of background noise like tool-calling conventions, metadata, "thinking", etc.
For small enough tasks with tight enough workflows, you can have it right now. Ie if you can constrain the task to work well with GPT OSS 120B/llama 3.3/qwen 3, then you can get upwards of 600 TPS on groq and up to 3k TPS on Cerebras.
Those models aren’t comparable to Opus, or even weaker models like MiniMax, but for certain task (focused context and prompts, strict workflows, single purpose requests) you absolutely can use these models and get insane speeds.
People seem to use these tools very differently from each other. I value intelligence over speed any day. My programs are written in Haskell, so there are rarely any tasks which require thousands and thousands of lines to solve. Just intelligence. If there are rote tasks, I want the LLM to help me find intelligent ways of automating it: the right abstraction, the right meta-programming technique.
I constantly push Opus and GPT, and they are getting better. But still have to do the hardest parts myself. I would not mind waiting 10-15 minutes for the right 20 lines of code!
Why do you use Haskell? Why not something that produces a more predictable memory use at runtime? (I’m asking earnestly as a former Haskeller turned Rustacean who sees the value in “Boring Haskell”, but favours strictness for anything internet-facing and many things that aren’t compilers.)
I use Haskell because purity and strong typing gives so much control over what each part of the program does. This has huge benefits when it comes to security, and just general lack of bugs. Also, it makes the code easy to write, once the types are in place.
I use Haskell because I find laziness to be a super power. I can solve so many problems in the most straightforward way, and then laziness saves my butt w.r.t. performance.
I use Haskell because it is a better C than C is. The foreign function interface is brilliant, and I can take C primitives and apply all the abstraction mechanisms from Haskell to them. My latest project has been OpenGL based, so lots of caring about byte alignments and shovelling data to the GPU. But all this can be automated with clever use of type classes and Generics (Haskells super cool meta system of data types.)
I use Haskell because I love applying abstractions to make code which describes the problem, and then the compiler finds the solution.
I don’t do programming for embedded, so I am rarely memory constrained. I also understand Haskell memory usage quite well, and can get myself out of trouble.
Google's 3.5 Flash – which came out yesterday – is 200-300 tokens/second (albeit purportedly inefficient in its use of reasoning tokens) and according to Google, 800-1500+ tokens/second on their 8i TPUs when they're out!
It's... suboptimal, but hopefully that's a reason to hope... if Google get themselves together for 3.5 Pro / the next Flash.
Why is that? It seems the other direction? I want to be sure I can complete a task in a certain amount of wall clock time. If the tokens per second are slow, then I am risking more by running a single approach at a time, and then have an incentive to try to multiplex my attention between separate work-streams. If the generation is fast enough to occupy my attention then there is no more available improvement by having parallel threads.
Obviously things will get expensive quick, but the main thing for me would be not dealing with the context switch every time I leave the agent to do stuff on it's own.
Feedback loops for prototyping could become even quicker.
Branching strategies, do 10 things in parallel and evaluate for the best at the end or something along the lines of an evolutionary algorithms. Turn up the temperature on an LLM and have a survival mechanism, and generate solutions to the same problem over and over.
Used them for a while! They didn't seem to have prompt caching so I burnt through the daily 24M token limitations really quickly when doing large scale changes on a codebase (essentially a team's worth of menial migration/refactoring work). A lot of it was okay, but plenty had to be re-done and I still spotted some issues months down the line, in part I blame their model catalogue which did get an update to GLM 4.7 sometime way back, but definitely is showing its age: https://inference-docs.cerebras.ai/models/overview
Quality wise, Anthropic gives me the best results (Opus for almost everything, I make sub-agents with fresh context review its work, after 2-10 loops, usually finds most issues). Token amount wise for agentic work, DeepSeek V4 is up there. What Cerebras is doing pretty cool though, apparently they even have prompt caching now like the other big providers: https://inference-docs.cerebras.ai/capabilities/prompt-cachi... At the same time, producing bad code faster was annoying in a uniquely new way.
Wish they'd update the models with their subscription, it could genuinely be great with the proper harness. Like if they can run GLM 4.7, surely they could at least get DeepSeek V4 Flash with a big context window going as a starting point. How can you have so much money to make your own chips, but can't run modern models that you can get for free? It's like they don't want people to use their subscription.
It’s GLM 4.7, GPT OSS 120B, or llama 3.1 8B so not exactly the latest or best models.
But GLM is good enough for many small tasks, certainly enough to get a taste for Cerebras’ high speeds!
[edit: actually that’s just their general models, I can’t see what Cerebras code offers. It was Qwen-coder when it launched but I don’t know what it is now. I think GLM 4.7 but I’m not completely sure]
> It was Qwen-coder when it launched but I don’t know what it is now.
This was also what I used at the time, the Qwen 3 Coder 480b on Cerebras. Worked great and was so stupidly fast it made me realize that if the hardware can be at that level and commercially available (say in a 5~10 years), for that price, then we will have entirely new bottlenecks. Human review at the pace it was going is completely impossible.
I just looked up what my computer is capable of (m2 MacBook Air) and it says 15-35 tokens per second. I could live with that writing code with a local model.
The sweet spot for being just fast enough to not irritate you is 10tok/s. Still slow but faster than you can sustain at typing and thinking. Just interesting to observe.
I think your demo need more realistic thinking logs because thinking usually burns at least 2x to 3x of tokens of the code and for harder tasks much more.
Indeed, at 30tok/s make it pause for 20 seconds while "thinking" is streaming (and hidden); that's the real experience.
It's interesting how even 5 tok/s is still much faster than you'd typically type, but feels glacially slow for an agent.
On the other hand, I've been using Mimo and Minimax a lot recently. They routinely reach 100-150 tokens per second and that feels too fast, to the point where it's hard to keep up with what it's actually doing. Great for subagents though.
They routinely reach 100-150 tokens per second and that feels too fast, to the point where it's hard to keep up with what it's actually doing.
There is no way you can follow what is going on even at 30 tokens per second. Maybe you can maintain a rough idea of what is going on for some tens of seconds but that is probably about it. Follow it in any detail, no chance. Reason about what you read, absolutely no chance.
800 tok/s — Cerebras-class, where the bottleneck is your eyeballs
I do not understand why they say this. I am not sure if it is even true, 800 tokens sounds like a page of text and I would assume you can look at one page per second without hitting any limitation of your eyes. Or is the resolution of the human not good enough to see an entire page at once and you have to scan it with the fovea? Scrolling text might of course hit the temporal resolution limit. But why does this even matter, your brain can not process anything close to the amount of information your eyes can take in.
Curious about the other way around, how many tokens per second a productive developer codes in a day?
Much more, given that you need to incorporate the dev's thought process too.
We truly are in the dial up era of GenAI.
This is awesome!! I use Cursor and I've been trending towards medium thinking models as much as possible - I don't like the dev cadence with something like opus 4.7 (thinking: very high) (great for some tasks, like complex plans). Eventually I'd like to make my way to open models and open harness, and this tool or something like it could help me understand what performance I'd need for productive work - bookmarked!
Cool visualization, but most of the token generation in my sessions doesn't go to output code or even the text I see. Reasoning tokens make up most of the output. That can only occur after processing the input files and context.
For non-trivial work I go through hundreds of thousands of tokens (combined prefill + tg of course) before even getting to some useful text output.
I mostly use LLMs for exploration and studies, rarely code generation. Prefill matters heavily for this. Even in the high hundreds or low thousands prefill rate I spend a lot of time waiting on the LLM (doing other things, not twiddling thumbs)
I'm flashing back to using a 1200 baud modem when the world was on 28.8k. Modems are much more regular-looking, though, since each character is a character. Unless you count color changes and such, which you only really notice at 1200...
Token/sec only makes sense once you tell me three four things:
1. decoding t/s, that is, when the model is generating text in the autoregressive fashion.
2. prefill t/s, that is, prompt processing speed.
3. What is the slope of those two numbers as the context size increases. An implementation that decodes at 50t/s with 2k context but decodes at 7t/s at 100k context is going to be a lot less useful that it seems at a first glance for a big number of real world use cases.
4. What's your use case? Reading a huge text and then having a small output like, fraud probability=12%? Or Reading a small question and generating a lot of text? This changes substantially if a model is usable based on its prefill/decoding speed.
For instance my DS4F inference on the DGX Spark does prefill at 350 t/s and at 200 t/s on already large contexts. But decodes at 13 t/s.
On the Mac Ultra the prefill is like 400 t/s and decoding 35 t/s.
The two systems can perform dramatically differently or almost the same based on the use case. In general for local inference to be acceptable, even if slow, you want at least 100 t/s prefill, at least 10 t/s generation. To be ok-ish from 200 to 400 t/s prefill, 15-25 t/s generation. To be a wonderful experience thousands of t/s prefill, 100 t/s generation.
Agreed. Prefill kills me for local model work. The model reads much faster than it writes, but I'd love to get a sense for how fast the model can read large source conversations.
This reminds me of when I signed up for cerebras to try it out and dumped $20 in and hooked it into opencode and the speed was truly insane. But my one session burnt through like $15 of that in seemingly a matter of minutes. I've since used those really high tok/s options for specific application use cases, but would not advise as a coding agent. Much harder to catch issues when it is moving a million miles an hour and then it is too late and it has already spent a ton of tokens.
Interesting. It seems to me that with that speed (20-30) on local hardware the real issue is quality of output, not tokens per sec.
It really depends. With the new "thinking" models they usually spend some time before writing the final answer. If they "think" for 1k tokens, that's a minute of spinning wheel you're gonna see for each question. Add that to the prompt processing, and diminishing speeds as context increases, and it becomes really slow for longer sessions.
Reminds me of the possibility of running DeepSeek at 3-4 t/s with SSD streaming, could be viable if you are running something overnight for example
The nice thing about DeepSeek and off-memory streaming is that you ought to be able to batch multiple sessions of it in parallel. Each individual session would slow down from streaming incrementallly more active weights from disk, but your total tok/s would ultimately only be limited by compute. Other models have trouble doing this, because the KV cache takes too much space in RAM (and increases wear-and-tear if stored on disk) even for somewhat limited context.
30tok/s looks fine when you're just streaming code, but the issue is that there's a lot of background noise like tool-calling conventions, metadata, "thinking", etc.
One thing I noticed is that prior to AI coding agents, I used to be able to tolerate 10 tokens/s. With AI Agents, I think 60 is minimum.
This is great. Agentic coding at 600+ tokens/sec is going to be a radically different beast. Coming soon-ish?
For small enough tasks with tight enough workflows, you can have it right now. Ie if you can constrain the task to work well with GPT OSS 120B/llama 3.3/qwen 3, then you can get upwards of 600 TPS on groq and up to 3k TPS on Cerebras.
Those models aren’t comparable to Opus, or even weaker models like MiniMax, but for certain task (focused context and prompts, strict workflows, single purpose requests) you absolutely can use these models and get insane speeds.
People seem to use these tools very differently from each other. I value intelligence over speed any day. My programs are written in Haskell, so there are rarely any tasks which require thousands and thousands of lines to solve. Just intelligence. If there are rote tasks, I want the LLM to help me find intelligent ways of automating it: the right abstraction, the right meta-programming technique.
I constantly push Opus and GPT, and they are getting better. But still have to do the hardest parts myself. I would not mind waiting 10-15 minutes for the right 20 lines of code!
Why do you use Haskell? Why not something that produces a more predictable memory use at runtime? (I’m asking earnestly as a former Haskeller turned Rustacean who sees the value in “Boring Haskell”, but favours strictness for anything internet-facing and many things that aren’t compilers.)
I use Haskell because purity and strong typing gives so much control over what each part of the program does. This has huge benefits when it comes to security, and just general lack of bugs. Also, it makes the code easy to write, once the types are in place.
I use Haskell because I find laziness to be a super power. I can solve so many problems in the most straightforward way, and then laziness saves my butt w.r.t. performance.
I use Haskell because it is a better C than C is. The foreign function interface is brilliant, and I can take C primitives and apply all the abstraction mechanisms from Haskell to them. My latest project has been OpenGL based, so lots of caring about byte alignments and shovelling data to the GPU. But all this can be automated with clever use of type classes and Generics (Haskells super cool meta system of data types.)
I use Haskell because I love applying abstractions to make code which describes the problem, and then the compiler finds the solution.
I don’t do programming for embedded, so I am rarely memory constrained. I also understand Haskell memory usage quite well, and can get myself out of trouble.
Google's 3.5 Flash – which came out yesterday – is 200-300 tokens/second (albeit purportedly inefficient in its use of reasoning tokens) and according to Google, 800-1500+ tokens/second on their 8i TPUs when they're out!
It's... suboptimal, but hopefully that's a reason to hope... if Google get themselves together for 3.5 Pro / the next Flash.
i really want a qwen on one of these chips: https://chatjimmy.ai
15k tokens/s would get me feeling like its actually worth splitting out worktrees to try several approaches to a problem
Why is that? It seems the other direction? I want to be sure I can complete a task in a certain amount of wall clock time. If the tokens per second are slow, then I am risking more by running a single approach at a time, and then have an incentive to try to multiplex my attention between separate work-streams. If the generation is fast enough to occupy my attention then there is no more available improvement by having parallel threads.
Do you have ideas/suggestions for agentic workflows that only start making sense at such speeds?
Obviously things will get expensive quick, but the main thing for me would be not dealing with the context switch every time I leave the agent to do stuff on it's own.
Feedback loops for prototyping could become even quicker.
Branching strategies, do 10 things in parallel and evaluate for the best at the end or something along the lines of an evolutionary algorithms. Turn up the temperature on an LLM and have a survival mechanism, and generate solutions to the same problem over and over.
If you have a Cerebras Code subscription you can experience it right now. Indeed, a very different experience.
Used them for a while! They didn't seem to have prompt caching so I burnt through the daily 24M token limitations really quickly when doing large scale changes on a codebase (essentially a team's worth of menial migration/refactoring work). A lot of it was okay, but plenty had to be re-done and I still spotted some issues months down the line, in part I blame their model catalogue which did get an update to GLM 4.7 sometime way back, but definitely is showing its age: https://inference-docs.cerebras.ai/models/overview
Quality wise, Anthropic gives me the best results (Opus for almost everything, I make sub-agents with fresh context review its work, after 2-10 loops, usually finds most issues). Token amount wise for agentic work, DeepSeek V4 is up there. What Cerebras is doing pretty cool though, apparently they even have prompt caching now like the other big providers: https://inference-docs.cerebras.ai/capabilities/prompt-cachi... At the same time, producing bad code faster was annoying in a uniquely new way.
Wish they'd update the models with their subscription, it could genuinely be great with the proper harness. Like if they can run GLM 4.7, surely they could at least get DeepSeek V4 Flash with a big context window going as a starting point. How can you have so much money to make your own chips, but can't run modern models that you can get for free? It's like they don't want people to use their subscription.
Have you tried Codex? If you have, how does it compare to Opus?
It’s GLM 4.7, GPT OSS 120B, or llama 3.1 8B so not exactly the latest or best models.
But GLM is good enough for many small tasks, certainly enough to get a taste for Cerebras’ high speeds!
[edit: actually that’s just their general models, I can’t see what Cerebras code offers. It was Qwen-coder when it launched but I don’t know what it is now. I think GLM 4.7 but I’m not completely sure]
> It was Qwen-coder when it launched but I don’t know what it is now.
This was also what I used at the time, the Qwen 3 Coder 480b on Cerebras. Worked great and was so stupidly fast it made me realize that if the hardware can be at that level and commercially available (say in a 5~10 years), for that price, then we will have entirely new bottlenecks. Human review at the pace it was going is completely impossible.
Nice, I always thought 15 tok/sec is too "slow"
Neat website, the visualization is great. I had a hard time wrapping my head around the tokens/s thing but this made it easy.
I just looked up what my computer is capable of (m2 MacBook Air) and it says 15-35 tokens per second. I could live with that writing code with a local model.
The non-linear scaling on the slider is an excellent UX.
On avg 1 token = 4 chars
So 75 tokens/s is ~ 300 chars per second which is the speed you'd get with a 2400 baud modem
Neat visual. 5 tok/s is still faster than me!
yeah 3t/s seems human. only that i never wrote code perfectly top to bottom.
I had the opposite reaction, 5tok/s is so slow that when you include all the reasoning and thinking + warmup it is far slower than me.
The sweet spot for being just fast enough to not irritate you is 10tok/s. Still slow but faster than you can sustain at typing and thinking. Just interesting to observe.
> Now switch between c and t at the same rate. The difference is striking — and intentional.
I don't see a big difference.
Thank you for this great utility. I love the "gut feel" calibration utilities like this one!
This is cool, thanks for making it.
super cool, thanks
This is great.