I wonder how proprietary it all is though, since the BitNet b1.58 paper has been out for a couple years now: https://arxiv.org/abs/2402.17764
From the wikipedia on 1.58 bit llms: "BitNet derives its performance from being trained natively in 1.58 bit instead of being quantized from a full-precision model after training. Still, training is an expensive process, and it would be desirable to be able to somehow convert an existing model to 1.58 bits. In 2024, HuggingFace reported a way to gradually ramp up the 1.58-bit quantization in fine-tuning an existing model down to 1.58 bits."
I just wonder how many of these labs are basically following the huggingface recipe here and possibly tweaking it and releasing models without huge training costs.
Independent testing of prismml suggest quite a capability drop off outside of their cherry picked benchmarks. I'll be curious to see what this model achieves though.
"Neutrino-1 8B was trained natively in its shipping format. There is no full-precision product model that was rounded afterward: the ternary representation is the medium the weights learned in, and the training methods that hold this quality at this depth are the lab’s unpublished work. The findings below are the part that travels."
This statement seems misleading at best.
Both the model page and the release page are basically unintelligible - I don't have a ton of faith in the work here, at least PrismML write coherent releases for their models.
Edit: Another beautiful piece of prose here, I almost wonder if they used the 8b model to generate the content for this release...
"Across the 6.95B coded weights, 62.63% sit at zero and the remainder splits 18.68% plus to 18.69% minus: sign-balanced to a hundredth of a point with no constraint asking for it."
> sign-balanced to a hundredth of a point with no constraint asking for it."
This isn't their model, this is (probably?) ChatGPT doing a brag / promo deck authorial voice. It routinely uses half a dozen sentence constructions that are relatively uncommon in normal or technical speech. Eccentric. Persuasive. Trying too hard. Restating its point in a promotional way that doesn't sound natural, leading into a sentence where it hyperbolically sells you on having done the impossible.
Normal persuasive speech uses these constructions, especially public speaking doing a VC pitch or an Ancient Aliens, but it would be fucking strange if a person started chaining them in normal conversation and using little else, trying to insinuate competence.
You start to recognize it pretty quickly on Youtube.
I really had high hopes for the larger Ternary Bonsai and it feels like there is scope to improve, but I get the sense (albeit a naïve, probably not fully informed sense) that improvement can perhaps only come by training directly into ternary.
I’ve actually been really impressed with the 27b model they recently released - amazing performance approaching 40 tok/s on m4 max and I didn’t run into any quality issues in the small set of tasks I tried. Haven’t gone full coding with it yet but suspect it’s better than say a 9b or 12b model.
I ran into some issues that are more extreme versions of Qwen's thinking loops while reasoning. It is great at the small puzzles I set for it but it did some frankly insane things on a PHP coding task I set it.
It also had some issues that might be parsing/chat template stuff, tool calling oddities. I will try it again, I did try it pretty much the day it shipped and it's possible there are more improvements in their llama.cpp fork since.
It would be churlish to be overcritical, mind you — the PrismML ternary stuff is an advance — but it feels like it should be applied at training. I figure we will see that, somewhere, quite soon.
Did you try the BottleCap ThinkingCap Qwen post-train with the reduced thinking overhead?
Not tested much but it is not noticeably worse than the underlying Qwen 3.6 27B in Q4_K_M, which in my experience is kind of a first for a Qwen fine-tune of this nature. They are almost always worse.
I think it does use fewer tokens while reasoning, which is potentially useful. I need to do more testing, because any performance advantage over the 27B is useful for me on an M1 Max.
Sorry I should have clarified - I meant that a ternary 27b model would outperform a non-quantized or 8 bit quantized 9 or 12b model - which it is generally close to (or much smaller than) in size. So yeah the comparison I was trying to make was between models of equivalent size or models that could run on similarly sized hardware.
Ah yeah, that makes a ton more sense :) I mean, what you said earlier also makes sense but was too obvious, now it makes sufficient sense, thanks for explaining :)
AI doesn't want anything, so it doesn't care whether it conveys meaning in its writing. And, apparently the developers of this project also don't care whether it conveys meaning. They just assume we'll wade through the slop? I dunno.
hey guys, founder here. I am working completely solo right now running on grant funding, so things slipped through on the website quality, it is largely written by AI, I'll deslopify the blogs and be much clearer about methodology. Neutrino-1 is a starting point for Fermion, It's a density experiment on how much capability survives per byte. rn working on really fast and efficient TTS and realtime models to which this architecture is generalizing pretty well.
Sorry for all the confusion from reading through the blogs, should be fixed in a couple of days!
There is not a single person mentioned on the website, github created 3 days ago, no real contact, everything hidden. Completely anonymous. Domain owner hidden.
Can’t say I’m a fan of containers for this. A big chunk of local LLM gains come (imo) from the open modular nature of llama.cpp and friends. Easy to modify. Easy to experiment.
Containers are the proprietary binary blob in hardware world equivalent
What? How are those even related with each other? You can just as easy modify and experiment with llama.cpp in a container as outside of it, they really shouldn't impact one another. Containers don't suddenly make llama.cpp less "open modular" somehow, and I'm not sure how you'd arrive as such conclusion.
are you pn linux? reproduced the same thing just now its pip dumping torch into /tmp while it installs, and on most distros /tmp is ram bound and capped at half your memory, so it fills up while your actual disk jsut sits there. export TMPDIR=/some/big/path before installing should fix it. adding a proper check next release so it actually tells you which folder ran out.
I just tune out. It’s not worth knowing, following every development in the field. If something works now it will probably work in 8 months even if it’s no longer the new hype thing. Who cares.
Not using any of it is also a valid option though it doesn’t satisfy your FOMO. But nothing ever will.
There's no need to, the 50 foot view is simply that many alternatives exist and they mostly fall into 3 meaningful weight classes with comparable performance among each class's members: too expensive to use indiscriminately, too big to run at home, and too small for complex work. As for names and faces in between, the overarching conclusion is that we're rapidly approaching commodity status and those don't really matter much
I'm working on solidifying local LLM work flow so I can stop tweaking the nobs and build things that arn't AI related.
There's tons of people caught up in the AI gold rush, but the value of real products for real people is out there. Even more so when you exclude the "I slapped AI onto this"
Can Dang implement a slop rating on submitted pages? Not a block but at least a % likelihood of AI slop content and that could also be tied with a BS rating as well.
Could use AI for both which seems hilariously appropriate.
There's a really interesting trend of labs using "proprietary" methods to convert existing models to a compressed ternary format.
PrismML actually targeted the same Qwen 8b model and got it down to 1.75gb here: https://prismml.com/news/ternary-bonsai
I wonder how proprietary it all is though, since the BitNet b1.58 paper has been out for a couple years now: https://arxiv.org/abs/2402.17764
From the wikipedia on 1.58 bit llms: "BitNet derives its performance from being trained natively in 1.58 bit instead of being quantized from a full-precision model after training. Still, training is an expensive process, and it would be desirable to be able to somehow convert an existing model to 1.58 bits. In 2024, HuggingFace reported a way to gradually ramp up the 1.58-bit quantization in fine-tuning an existing model down to 1.58 bits."
The section from huggingface is here: https://huggingface.co/blog/1_58_llm_extreme_quantization#fi...
I just wonder how many of these labs are basically following the huggingface recipe here and possibly tweaking it and releasing models without huge training costs.
Independent testing of prismml suggest quite a capability drop off outside of their cherry picked benchmarks. I'll be curious to see what this model achieves though.
Unfortunately Fermion Research appears to entirely AI generate all of their content here, even for the research section: https://www.fermionresearch.com/research/neutrino-8b/
"Neutrino-1 8B was trained natively in its shipping format. There is no full-precision product model that was rounded afterward: the ternary representation is the medium the weights learned in, and the training methods that hold this quality at this depth are the lab’s unpublished work. The findings below are the part that travels."
This statement seems misleading at best.
Both the model page and the release page are basically unintelligible - I don't have a ton of faith in the work here, at least PrismML write coherent releases for their models.
Edit: Another beautiful piece of prose here, I almost wonder if they used the 8b model to generate the content for this release...
"Across the 6.95B coded weights, 62.63% sit at zero and the remainder splits 18.68% plus to 18.69% minus: sign-balanced to a hundredth of a point with no constraint asking for it."
I guess it's saying how many of the weights are -1, 0 or +1.
It is, but why? And what's with the bizarre way of phrasing that? Why the bizarre observation that, indeed, nobody asked for it?
> sign-balanced to a hundredth of a point with no constraint asking for it."
This isn't their model, this is (probably?) ChatGPT doing a brag / promo deck authorial voice. It routinely uses half a dozen sentence constructions that are relatively uncommon in normal or technical speech. Eccentric. Persuasive. Trying too hard. Restating its point in a promotional way that doesn't sound natural, leading into a sentence where it hyperbolically sells you on having done the impossible.
Normal persuasive speech uses these constructions, especially public speaking doing a VC pitch or an Ancient Aliens, but it would be fucking strange if a person started chaining them in normal conversation and using little else, trying to insinuate competence.
You start to recognize it pretty quickly on Youtube.
I’m hoping it becomes a trend for mocking AI output by ironically using AI tells.
Also, wouldn't normalisation tend to produce a mean of zero in the weights, anyhow?
I really had high hopes for the larger Ternary Bonsai and it feels like there is scope to improve, but I get the sense (albeit a naïve, probably not fully informed sense) that improvement can perhaps only come by training directly into ternary.
I’ve actually been really impressed with the 27b model they recently released - amazing performance approaching 40 tok/s on m4 max and I didn’t run into any quality issues in the small set of tasks I tried. Haven’t gone full coding with it yet but suspect it’s better than say a 9b or 12b model.
I ran into some issues that are more extreme versions of Qwen's thinking loops while reasoning. It is great at the small puzzles I set for it but it did some frankly insane things on a PHP coding task I set it.
It also had some issues that might be parsing/chat template stuff, tool calling oddities. I will try it again, I did try it pretty much the day it shipped and it's possible there are more improvements in their llama.cpp fork since.
It would be churlish to be overcritical, mind you — the PrismML ternary stuff is an advance — but it feels like it should be applied at training. I figure we will see that, somewhere, quite soon.
Did you try the BottleCap ThinkingCap Qwen post-train with the reduced thinking overhead?
I haven't spun up thinkingcap yet but I'm aware of it and am intending to try it out soon. How did you find it?
Not tested much but it is not noticeably worse than the underlying Qwen 3.6 27B in Q4_K_M, which in my experience is kind of a first for a Qwen fine-tune of this nature. They are almost always worse.
I think it does use fewer tokens while reasoning, which is potentially useful. I need to do more testing, because any performance advantage over the 27B is useful for me on an M1 Max.
> suspect it’s better than say a 9b or 12b model
Whaaat, a 27b model might be better than 9b or 12b model? What would make you do such an outrageous claim?
Sorry I should have clarified - I meant that a ternary 27b model would outperform a non-quantized or 8 bit quantized 9 or 12b model - which it is generally close to (or much smaller than) in size. So yeah the comparison I was trying to make was between models of equivalent size or models that could run on similarly sized hardware.
Ah yeah, that makes a ton more sense :) I mean, what you said earlier also makes sense but was too obvious, now it makes sufficient sense, thanks for explaining :)
All you need is Ternary Aware Training and for AI researchers to come up with a backronym for TIT.
Wouldn’t that be TAT?
The content on that page is too AI-generated to make sense to me; I don't understand what the model is for.
AI doesn't want anything, so it doesn't care whether it conveys meaning in its writing. And, apparently the developers of this project also don't care whether it conveys meaning. They just assume we'll wade through the slop? I dunno.
It has no ability to ground output in causality. It does not understand what causes humans to learn. It’s a correlation and randomness engine.
I have the same question.
I get that it’s designed to run on a CPU, big GPU or MacBook (although the way that was phrased confused me at first).
I’m struggling with what a “decoder-only” model is good for.
hey guys, founder here. I am working completely solo right now running on grant funding, so things slipped through on the website quality, it is largely written by AI, I'll deslopify the blogs and be much clearer about methodology. Neutrino-1 is a starting point for Fermion, It's a density experiment on how much capability survives per byte. rn working on really fast and efficient TTS and realtime models to which this architecture is generalizing pretty well.
Sorry for all the confusion from reading through the blogs, should be fixed in a couple of days!
Lots of respect for a solo founder. But you submitted this yourself, and it’s incomprehensible.
For me, sloppy content is worse than no content. I'd rather have 3 sentences written by a human than a wall of Claudespeak.
Do you have an RSS or a way to subscribe to the TTS work you are doing? Your newsletter button is just a mailto link
There is not a single person mentioned on the website, github created 3 days ago, no real contact, everything hidden. Completely anonymous. Domain owner hidden.
Largely outperformed by Ternary-Bonsai-8B by their own chart, doesn't seem clear what their special sauce is here.
Can’t say I’m a fan of containers for this. A big chunk of local LLM gains come (imo) from the open modular nature of llama.cpp and friends. Easy to modify. Easy to experiment.
Containers are the proprietary binary blob in hardware world equivalent
What? How are those even related with each other? You can just as easy modify and experiment with llama.cpp in a container as outside of it, they really shouldn't impact one another. Containers don't suddenly make llama.cpp less "open modular" somehow, and I'm not sure how you'd arrive as such conclusion.
I was really hoping this would have something to do with neutrinos
Damn, I was hoping this was about an actual neutrino.
software naming claims another victim
So... not a new subatomic particle discovery then.
So I'd like to see a Nemotron3 Ultra converted to a 1.58bit format then have them retrain on the open dataset.
Unfortunately it crashed out 'no space left on device' while installing the python demo/quickstart.
Only problem was there is plenty of space on the device. PLENTY (not quite 750gb).
are you pn linux? reproduced the same thing just now its pip dumping torch into /tmp while it installs, and on most distros /tmp is ram bound and capped at half your memory, so it fills up while your actual disk jsut sits there. export TMPDIR=/some/big/path before installing should fix it. adding a proper check next release so it actually tells you which folder ran out.
There’s a new announcement every other day wrt models. How do y’all keep track of them all same know what’s decent? Good grief!
And if it’s decent today, it’s shit in eight months! I tool hop as much as the next dev but this is a bit much.
I just tune out. It’s not worth knowing, following every development in the field. If something works now it will probably work in 8 months even if it’s no longer the new hype thing. Who cares.
Not using any of it is also a valid option though it doesn’t satisfy your FOMO. But nothing ever will.
There's no need to, the 50 foot view is simply that many alternatives exist and they mostly fall into 3 meaningful weight classes with comparable performance among each class's members: too expensive to use indiscriminately, too big to run at home, and too small for complex work. As for names and faces in between, the overarching conclusion is that we're rapidly approaching commodity status and those don't really matter much
I'm working on solidifying local LLM work flow so I can stop tweaking the nobs and build things that arn't AI related.
There's tons of people caught up in the AI gold rush, but the value of real products for real people is out there. Even more so when you exclude the "I slapped AI onto this"
The real strength of an 8B model is efficiency. It will be interesting to see the balance between performance and inference cost.
This all sounds like middle-out
AI slop site with AI slop research...
Blog populated with incoherent PR material generated by Yet Another AI.
Sigh...
Can Dang implement a slop rating on submitted pages? Not a block but at least a % likelihood of AI slop content and that could also be tied with a BS rating as well.
Could use AI for both which seems hilariously appropriate.
Or perhaps a slop flag on submissions that gets added once enough readers have flagged it as such. I'd love that feature.
Slop article, slop site... slop model?
[flagged]
[flagged]