When an LLM thinks, it typically just makes one pass. It outputs tokens from top to bottom, beginning to end, and then it's done. But when people think, especially strong thinkers, we typically iterate and revise our thoughts on the fly. We do numerous passes. We stop and restart, we reconsider, we review, we reevaluate. Sometimes we do this so quickly and automatically that we don't even realize we're doing it. I think a lot of what separates a highly intelligent or effective person from others has less to do with the quality of their first pass and more to do with just how many additional passes they're able to do in the same amount of time, and of course what kind of criteria they're habituated to consider during their review passes.
Introspecting about this is difficult, but experimenting with LLMs is easy. First, simply ask an LLM to do something complex. For example, to come up with a new business idea, or to plan the next month of your life, etc. After it finishes, tell it:
"Review what you just wrote, according to some appropriate list of evaluation criteria that you come up with first. And then, based on the results, iterate and generate a better response if warranted."
It's insane how much better the next answer will usually to be. Often it'll catch and erase tons of hallucinations, logical errors, and inefficiencies. And you can simply copy-paste this again and again until you begin to hit diminishing returns. Or, in a harness like Claude Code, for example, I might shortcut this whole process by saying, "Use sub-agents to iteratively review and iterate on your work until convergence."
The reason why most people don't prompt LLMs to do this (besides simply not thinking of it) is that it takes time.
But what if it didn't?
What if the LLM's response came back in milliseconds rather than minutes? Then there would be almost no reason NOT to do this. In fact, one could almost imagine it baked into the assistant/harness -- a massive step change in practical quality, enabled by nothing more than speed.
Similarly, if I ever get a bit too vibey and don't carefully review code changes myself, the blast radius is generally significantly resolved by a carefully tuned "did you consider x, y, and z" skill after a first draft partnered with a "deploy an adversarial review agent for the worktree".
Turns out TDD is far better for robots than humans, who knew
Neither the Cerebras or OpenAI post [0] outright state that this performs exactly the same as regular 5.6 Sol. I feel if this was 1:1 just Sol but much faster, they'd (rightfully) scream that off the rooftops. A line such as "this is the same performance, just faster, with no downsides" would go a long way in clarity and communication. Along with no pricing information, I'll hold out on further information.
Same for massive performance differences in the way providers like Cerebras, Groq, etc. have deployed models including K2.6 on Cereberas specifically. Massive deltas in tool call and overall quality despite there being far more clarity in open weight vs proprietary model deployment.
The AA suite graph with that animation is the only time in either post that absolute parity is being asserted and I'd be amazed if that was the case, but am doubtful why their phrasing is so cagey.
Why not assert full parity in writing? It "performs the same (within run-to-run variance) across all evals that Sol has been tested with" is very different to "no quality compromise/degradation", the later allowing for a lot more wiggle room and interpretation in what evals you use to assess that, what quality truly means, etc., the former meaning identical in all situations.
Could also be a language barrier here in fairness, maybe this phrasing is more iron clad than I give them credit, but especially with OpenAI, I have seen enough checkpoints asserted as unchanged in "quality" to where I am skeptical. Ironically, I never saw that with Anthropic (which has gotten far more heat for degradation accusations) while a model was deployed with one exception in mid-late April this year. Pure speculation, but believe it wasn't noticed much before "agentic coding" became more popular, because chat output is far more subjective without a rating framework vs code passing which can be an objective metric with more potential for frustration.
> It’s also our best model for many non-chat use cases—we’ve seen early testers migrate from text-davinci-003 to gpt-3.5-turbo with only a small amount of adjustment needed to their prompts.
That's why I still remember this so well, they claimed one model to be their best and a straight up drop-in during deprecation when in my (back then even more amateurish then today) testing this was plainly not the case. A model cannot be "best" if it's measurably worse in many situations, then what was still available at the time.
[0] https://openai.com/index/gpt-4-api-general-availability/
[1] https://openai.com/index/introducing-chatgpt-and-whisper-api...
There is no pricing info, which could mean it's "if you have to ask..." territory or they are simply gauging interest before deciding
Who knows if they will subsidizes it to mitigate sticker shock, but it's a safe assumption that it will be scarily expensive. However if you are in a "cost is no obstacle, speed is god" position, it will likely be pure magic.
I don't think I understand why they aren't leveraging the increased speed to do batching to serve more customers at a "normal" tok/s.
Is the limitation, even on cerberus, still that the cache can only serve so many concurrent sessions over time? Is there no scaling advantage? I genuinely do not understand how any of this works.
With the tool calls that can be done, you're not pricing this against an executive assistant or pocket analyst - you're pricing this against the ability to have an entire Bourne Identity style analysis room at your disposal. The limited inventory will go to the people for whom money is no object.
So, afaik, Cerebras are optimizing for ultra-low latency batch=1 inference.
https://newsletter.semianalysis.com/p/cerebras-faster-tokens... goes quite in-depth.
Also worth looking into how they do cooling for it, because that's kind of absurd and awesome as well.
There's some technical hypotheses about it that other people are offering.
But also from a business perspective, it totally makes sense not to go any sort of batching play. It's really valuable and very clear to consumers to make your pitch entirely about lower latency rather than higher bandwidth.
There are so many scenarios that are latency-constrained that will be difficult or even impossible for someone even with fleets of high-bandwidth compute to compete with you on.
Very easy pitch to sell a customer who asks what differentiates you from other companies: you pay us a premium for lower latency than anyone else.
Batching works because of severe memory bottleneck, but Cerebras whole thing is serving models out of "L1 cache" (?).
One way large models are served on a bunch of cerebras chips is by essentially distributing layers' weights across chips. Few layers's weights per chip - as many as the KV cache + activations + weights will allow. You use pipelining to hide the latency of the inter-chip 150 GB/s link.
On GPUs, you amortize the cost of loading weights from HBM to SRAM across multiple users - thereby making it cheaper _per_ user. But here, there is no such amortization. The weights are already there. It is the activations that stream through.
You _could_ do batching/continuous batching, but that would just service more users at lower token/s each without any amortization of fixed cost, due to fixed cost (loading weights) being non-existent.
Awesome work. I'm personally very excited for faster models/inference.
I think speed is underrated to some degree in the current conversation. For a while, I was using Cursor's Composer quite a lot, even over frontier models, just because of how darn fast it was.
IME waiting for an agent to work through a problem is a detriment to attention span; your mind drifts to other things while you wait. Maybe you can steer several agents in a round robin instead, but then there's a cognitive tax from context switching. Faster models mean fewer gaps in focus.
It also spent almost 800k tokens on these lines…
I suspect Humanity's Last Exam is without tool-calls, making it kind of the perfect benchmark to highlight how fast Ultrafast is, but not really the same as the everyday work you or I do.
Rught now on large scale codebases the bottleneck is both claude/codex inference, as well as time it takes to run tens of thousands of tests. We put those workloads on dedicated epyc 9005 build machines - but it still takes minutes per run. Those who can afford the fast tokens will be in the market for faster CPU that money can buy today.
i know people are joking about the sol ultrafast prices (its unlikely to be accessible for average joes) but this shows scaling wafer cores works for inference boost
which makes me very excited, sol ultrafast will be as slow as it will get if that makes sense. at these token speeds , we will see a much deeper economic impact.
It is a bit outdated (scores ± 40% lower), but smart enough for a lot of coding tasks, and can cost under 1/10th of Sol.
> delivering 17k tokens per second per user on Llama 3.1 8B model.
Obviously this is a much smaller model, but I really can't wait for ASICs to take over the LLM space.
Imagine running a model like Sol/Fable (even half the size with 60-70% of it's intelligence) on your own ASIC hardware.
Some interesting twitter analysis here:
My prediction is that, this time next year, top developers outside ai labs will be spending 50k USD+ on inference.
Within labs, I've heard spend is already far beyond this per developer.
I think it is more like top companies, not top developers, and the problem with developers in top companies was and is - absolute majority of them are not actually directly working on things that increase revenue, so companies can spend a ton of money and see barely if any changes in the product and the bottom line, so companies, at least legacy ones will be reluctant to sponsor that long term.
If someone subsidize maybe, but if the companies need to pay no way, unless there is hard evidence of the return.
Obviously there are a ton of ways to spend money / tokens and people have different levels of experience that will put this ceiling at very different levels for different people.
Given sufficient budget and scope, I could certainly productively burn a half million dollars in tokens a year or more. I think that's where we're headed anyway, buying a 2nd or 5th claude max subscription feels slightly excessive for personal usage, but at a corporate level...
And a moderately heavy user dev can easily spend a few $K a month, so yeah. Not impossible, but a high cost, and the diminishing returns definitely kick in
If it truly is only ~1-2T parameters, then this kinda kills 2 narratives for me.
1. all the handwringing about open source catching up via Kimi K3 (3T params) is complete nonsense. All that matters imo for determining which labs are leading is intelligence per parameter. Anyone with a enough compute can train a giant model, but being able to squeeze capabilities into smaller models gives you a massive inference and training edge.
2. Inference margins are clearly insane, and this explains why OpenAI was able to lower the price of Luna by 80%. Id guess that thing is probably 120b params based on the TPS they are serving it at.
I didn't like Sol initially but it is growing on me the more I use it. Its personality is a bit flat and I caught it taking shortcuts a few times. But once I learned how to interact with it, I'm genuinely warming up to it. I find that it writes code that has fewer bugs even than Fable (although, to be fair I reach for Fable when the task is less well defined).
If this has similar performance to Sol at max reasoning level, this would be a compelling reason to shift even more of my work (maybe the majority) to this model.
I think it's only a matter of time before miniaturization can have a thumbnail sized user-replaceable accessory that contains the LLM built onto the hardware. I admit I don't know how any of that works, but would be amazing to experience. Fully local, fully offline, ultra fast local inference better than any personal computing product.
High end phones can already run the smaller models at enough speed to be probably useful, especially for background/overnight photo tagging and curation and things like that.
But if humans need to check its work, then 10X speed doesn’t really matter I guess.
However, I wouldn't diss the "ultraspeed" options untill I try them. Having agent thinking become near instant could change the way I (or you) use agents.
A human could have an agent run 10x more correction checks. If even after that they still need to check manually for issues, then they really need to work on their specification skills.
Curious, what are some of the use cases?
Compilation time will be a genuine bottleneck for slop coding if this becomes the standard generation rate over the next few years. Go, Zig or even C99 with TCC for dev builds, any language that can get you systems-level performance (or close to it) in a dev environment where you can iterate in ms rather than minutes is going to be immensely more appealing than generating a potential prototype in 10 seconds and waiting 15 minutes for it to compile.
printf("Hello, world");
48 65 6C 6C 6F 2C 20 77 6F 72 6C 64
48 83 EC 28
48 8D 0D F5 0F 00 00
E8 F0 00 00 00
33 C0
48 83 C4 28
C3
Why? Because you are defining the implementation based on its observable behaviour rather than as a rule set to be followed.
> In our evaluations, GPT-5.6 Sol on Ultrafast mode answered all 2,500 HLE questions in 11 hours and 11 minutes. Claude Fable 5 needed 78 hours and 27 minutes, more than three days of continuous compute, to arrive at the same conclusions. In other words, Ultrafast worked through the frontier of human knowledge in a single working day, achieving comparable accuracy nearly 7× faster.
This is actually insane.
Hopefully the release ultrafast of Terra and Luna too.
In fact, I'd say it's overqualified for the kind of work I'm doing, because it spends >half the time verifying trivial changes (and the verification isn't as helpful as you'd expect, even with bigger models).
Maybe I can prompt it to be less aggressive about that (the new GPT models do it even without prompting).
Anyway, Ultrafast Luna would be amazing, though I strongly doubt they can offer Cerebras at anything approaching the current prices. Now we wait for Moore's Law? :)
So is calculating the total time required to answer all of the questions.
But why is that important if they're measuring time?
Don’t blink.
(Chatjimmy has 14,200 TPS.)
700 TPS with reasoning is awesome and it speeds things up.
Cerebras as public traded company is worth keeping an eye what they produce.
As the sibling comment to yours mentioned, if they had a reasoning model “hardware-ified” onto a custom chip (as is their plan for IIRC this or next year, a new ASIC), it’d output fast decode speeds for the regular output as well as reasoning sections. Both would be ≈equally fast.
I recognize that the former is the multiplication symbol, but I don't think it should be used that way.
[0] https://news.ycombinator.com/item?id=35490837