The pareto frontier needs clearer distinction. Benchmarks miss half the story. What, if any, capability is lost by the token reduction (for example, was it like super awesome at Golang before and now kind of sucks? that kind of distinction).
Ignoring for the moment issues of what "counts" as open, won't open models rapidly advance due to stuff like this in ways that it's less possible for the proprietary ones to do? This is exactly how Linux & Wikipedia, for example, overtook their "frontiers", right?
not necessarily, just knowing something is possible will motivate others to achieve it somehow. Which is why there are so many LLMs and OAI doesn't have a monopoly
Ignoring for the moment issues of what "counts" as open, won't open models rapidly advance due to stuff like this in ways that it's less possible for the proprietary ones to do? This is exactly how Linux & Wikipedia, for example, overtook their "frontiers", right?
I suspect the advantage that catapulted Linux ahead of the establishment was less technical potential and talent and more organizational advantage. That's not to diminish the technical talent of the Linux crew, but them being unencumbered gave them more degrees of freedom. The rest is history.
So as long as the AI companies don't succumb to "big company" dynamics, they can outlead. To wit: Open AI and Anthropic are kicking Google's ass.
Diff people have diff motives to experiment, then new work is done on top of stuff that "hits" in a way no one anticipated. Then work gets piled on top in a way that might make it hard to port
then new work is done on top of stuff that "hits" in a way no one anticipated.
Indeed. And when you have freedom to play, you are able to find new stepping stones that you didn't anticipate. And you can combine stepping stones in new ways to make new discoveries.
The difference between contributing to OS and AI, is that the first is a hobby alternative to woodworking or hiking, while the other can easily bootstrap you a company you can get millions in investment, at least for time being.
On the smaller end, Quen 3.8, while being extraordinarily capable for a small local model, also suffers from extreme thinking. I wonder if the techniques described here generalize to other models too.
I suspect it might generalize to other large models, but I don't think Qwen3.8 27B is one of them. Kimi K3 is a 2.8 trillion parameter model, and I suspect that is playing a big role in being able to reduce the length of CoT without taking a hit in quality.
Technically kimi k-3 weights license is not open weight (it has a lot of restrictions). I would classify it as ‘weight open’ similar to the bsl and fsl ’source open’ licenses.
There is little to no point reading the article as well. It's stripped of all alpha.
task and environment feedback
on-policy planning and learning
feedback connects decisions to their consequences
These are deliberately the least informative phrases you could possibly use to describe what you have done, while still being in the realm of words that go over a generic investor who has no idea whats going on and may be dazzled by sciencey sounding language.
Cursor compose 2.5 article where they used and described on policy self distilation was actual alpha.
Off topic:With sol pricing drop tbh kimi k3’s value prop has not been that great. For our internal use case/testing/benchmarks sol come out with way better quality and much cheaper costs.
Kimi really needs to drop their pricing (I heard it’s set by them across all the neoclouds)
Sol is at 2/10 vs kimi’s 3/15
Agreed, I think the only place where it’s still interesting is ui design. Visually kimi and muse feel much nicer than frontier models to me, but maybe it’s an artifact of everything terrible being Claude Design
This is really interesting. I think the Fireworks Serverless Training infrastructure they used to develop it is also unique and needed. Except if someone works at one of a handful of the largest labs, it is very difficult to set up or try any sort of training pipeline. The managed training infrastructure makes it available to more people.
This is partly the appeal of Jev et al; having a quick model for simple tasks, that doesn’t require that much thinking
It’s amazing all the workflows that models like that can unlock. And yes, classifiers and other ML models have been around for a while for these types of tasks, but Jev has made it easy and cheap to play and experiment. This in turn, is incentivizing people to try them for a bunch of stuff, unlocking creativity and producing a lot of new cool (and eventually potentially very useful) applications
Lots of use cases!
I've personally used it for the following:
1. Evals (once you have your rubric defined and tuned using a reasoning model, jev can be great for running periodic evals especially those that run daily.
2. e-commerce catalog classification
3. quick search using anything as context and query mapping to a pre-defined set.
The result? Ember-1 set a new Pareto frontier for Bedside Bench across both open and closed models including GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5 on cost/task.
Obviously this research was done before 6.0 Sol and Opus 5.5 came out. Your point stands that the frontier moves quickly and small gains can be eclipsed quickly.
This is the golden age of model training. Some days ago, I decided I wanted a local CPU only model that can perform exceptionally well for English to Bash translation (to avoid the googling for command syntax). I got a bunch of subagents to generate large amount of training data (140k+ samples), got the Qwen 3 0.6B base model, pointed Astra at it, and off to the races. It trained for 2 days (on and off) and I got a surprisingly good model for my task! The total active time I spent was a few hours. And it is still improving, what a time to be alive!
Analysis paralysis stifles not just human intelligence, but other intelligences too.
I don't think the article mentions Pareto frontier enough.
Also, did I miss a memo? Suddenly every article on AI seems to be talking about the Pareto frontier - or have I just not been paying attention?
I guess they figure "best bang for your buck" comes off a little too colloquial.
They want it to be the best at something. And it's obviously not the absolute smartest. So here we are.
Pareto frontier on some benchmark that I am hearing of for the first time.
Kimi K3 with less reasoning tokens isn't exactly exciting either, and particularly so if the license is less open than original Kimi K3.
Need this done for DeepSeek, ideally one of the Flash models.
If you have the compute, I have the expertise.
And GLM. Both Deepseek 4.1 Flash and GLM 5.3 Flash are quote verbose when thinking.
Well done, and great iteration.
The pareto frontier needs clearer distinction. Benchmarks miss half the story. What, if any, capability is lost by the token reduction (for example, was it like super awesome at Golang before and now kind of sucks? that kind of distinction).
Unfortunately, it’s hard to make a chart of that.
It looks like it would be similar to GLM 5.3 Flash, had they tested it...
Ignoring for the moment issues of what "counts" as open, won't open models rapidly advance due to stuff like this in ways that it's less possible for the proprietary ones to do? This is exactly how Linux & Wikipedia, for example, overtook their "frontiers", right?
Yes, absolutely, but only if people keep contributing in the open.
not necessarily, just knowing something is possible will motivate others to achieve it somehow. Which is why there are so many LLMs and OAI doesn't have a monopoly
No, because close labs/models borrow but don't contribute back.
I suspect the advantage that catapulted Linux ahead of the establishment was less technical potential and talent and more organizational advantage. That's not to diminish the technical talent of the Linux crew, but them being unencumbered gave them more degrees of freedom. The rest is history.
So as long as the AI companies don't succumb to "big company" dynamics, they can outlead. To wit: Open AI and Anthropic are kicking Google's ass.
Diff people have diff motives to experiment, then new work is done on top of stuff that "hits" in a way no one anticipated. Then work gets piled on top in a way that might make it hard to port
Indeed. And when you have freedom to play, you are able to find new stepping stones that you didn't anticipate. And you can combine stepping stones in new ways to make new discoveries.
Greatness cannot be planned.
The difference between contributing to OS and AI, is that the first is a hobby alternative to woodworking or hiking, while the other can easily bootstrap you a company you can get millions in investment, at least for time being.
Won't the "frontier" labs figure out whatever techniques were used and apply them to their closed models?
On the smaller end, Quen 3.8, while being extraordinarily capable for a small local model, also suffers from extreme thinking. I wonder if the techniques described here generalize to other models too.
I suspect it might generalize to other large models, but I don't think Qwen3.8 27B is one of them. Kimi K3 is a 2.8 trillion parameter model, and I suspect that is playing a big role in being able to reduce the length of CoT without taking a hit in quality.
That's just vibes, though.
Does anybody know if this would be a good model for creative writing?
So they trained a model on open weights, and then aren't releasing the weights... am I reading this right?
Technically kimi k-3 weights license is not open weight (it has a lot of restrictions). I would classify it as ‘weight open’ similar to the bsl and fsl ’source open’ licenses.
weight available
It happens. Most open licenses aren't GPL style copyleft.
It happens with open source software all the time, why would we expect any different with open source weights.
[delayed]
Aren't Cursor Composer models like this too? At some point all the extra RL you do can be considered as proprietary information added.
Not suggesting this is right or wrong, but is sort of the nature of the technology.
There is little to no point reading the article as well. It's stripped of all alpha.
These are deliberately the least informative phrases you could possibly use to describe what you have done, while still being in the realm of words that go over a generic investor who has no idea whats going on and may be dazzled by sciencey sounding language.
Cursor compose 2.5 article where they used and described on policy self distilation was actual alpha.
Off topic:With sol pricing drop tbh kimi k3’s value prop has not been that great. For our internal use case/testing/benchmarks sol come out with way better quality and much cheaper costs. Kimi really needs to drop their pricing (I heard it’s set by them across all the neoclouds) Sol is at 2/10 vs kimi’s 3/15
Agreed. Even on the open weight side, GLM 5.3 has roughly equivalent performance to Kimi K3 for less than half the cost.
Agreed, I think the only place where it’s still interesting is ui design. Visually kimi and muse feel much nicer than frontier models to me, but maybe it’s an artifact of everything terrible being Claude Design
This is really interesting. I think the Fireworks Serverless Training infrastructure they used to develop it is also unique and needed. Except if someone works at one of a handful of the largest labs, it is very difficult to set up or try any sort of training pipeline. The managed training infrastructure makes it available to more people.
I can’t help but think it’s more expensive tinker.
Do they mean Opus 5.5 or Opus 5?
This is partly the appeal of Jev et al; having a quick model for simple tasks, that doesn’t require that much thinking
It’s amazing all the workflows that models like that can unlock. And yes, classifiers and other ML models have been around for a while for these types of tasks, but Jev has made it easy and cheap to play and experiment. This in turn, is incentivizing people to try them for a bunch of stuff, unlocking creativity and producing a lot of new cool (and eventually potentially very useful) applications
What are the useful applications of Jev so far? Not to sound dismissive, I just haven’t seen what people are using it for yet.
Lots of use cases! I've personally used it for the following:
1. Evals (once you have your rubric defined and tuned using a reasoning model, jev can be great for running periodic evals especially those that run daily.
2. e-commerce catalog classification 3. quick search using anything as context and query mapping to a pre-defined set.
The result? Ember-1 set a new Pareto frontier for Bedside Bench across both open and closed models including GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5 on cost/task.
"Pareto": 8 hits
"Opus 5.5": zero hits
Obviously this research was done before 6.0 Sol and Opus 5.5 came out. Your point stands that the frontier moves quickly and small gains can be eclipsed quickly.
This is the golden age of model training. Some days ago, I decided I wanted a local CPU only model that can perform exceptionally well for English to Bash translation (to avoid the googling for command syntax). I got a bunch of subagents to generate large amount of training data (140k+ samples), got the Qwen 3 0.6B base model, pointed Astra at it, and off to the races. It trained for 2 days (on and off) and I got a surprisingly good model for my task! The total active time I spent was a few hours. And it is still improving, what a time to be alive!