Open access for next 5 hours (Ternary-Bonsai-8B-Q2_0.gguf, running on RTX 3090) or until server crashes or the this spot instance gets taken away :) =>
Fyi, I believe `--flash-attn on` doesn't do anything, you should instead use `--flash-attn 1`.
I'm getting ~150t/s on a RTX 3080 10GB as well with f16 cache type.
played around with it a little, works great for
- basic conversational tasks
- basic reasoning
- creating little UI snippets
and not good for one shot coding tasks.
once it starts hallucinating about something you can't make it rectify that. during a coding task it wrote some madeup library function, despite of pointing it out, giving docs, it kept producing the same code with the madeup function all while acknowledging the problem.
btw, thanks for putting it together!
Thanks a lot, I was about to clone their llama.cpp branch and do the same.
Some more interesting tidbits from my go-to tests:
* Fails the car wash test (basic logic seems to be weak in general)
* Fails simple watch face generation in html/css.
* Fails the "how many Rs in raspberry test" (not enough cross-token training data), but will funnily assume you may be talking about Indian Rupees and tell you a lot about raspberry prices in India without being asked. Possible Indian training data unbalance?
* Flat out refuses to talk about Tiananmen square when pushed directly - despite being from a US company. Again, perhaps they are exposed to some censored training data? Anyways, when slowly built up along the conversation by asking about locations and histories, it will eventually tell you about the massacre, so the censorship bias seems weak in general. Also has no problem immediately talking about anything Gaza/Israel/US or other sensitive topics.
* Happily tells you how to synthesize RDX with list of ingredients and chemical process step by step. At least it warns you that it is highly dangerous and legally controlled in the US.
Thanks, already suspected as much. Also gives context to the other comment here that says it is basically equivalent in accuracy to Qwen3.5-4B. Essentially seems to be a very good quantization of that model, not a new BitNet.
It's a good-per-byte-but-not-in-absolute-terms quantization of Qwen3-8B that's comparable in accuracy to Qwen3.5-4B at 4-bit quantization (which makes the 4B model larger in terms of storage, though the lower number of parameters and hybrid attention give it a speed advantage if you're not bottlenecked on memory bandwidth for the model weights.)
in my results, accuracy-wise Ternary-Bonsai-8B is on par with Qwen3.5-4B. But in accuracy-per-byte, bonsai is the clear winner:
=> Ternary-Bonsai-1.7B achieved 65.1% from 462 MiB, beating Qwen3.5-0.8B by 12 points while being ~5% smaller on disk.
=> Ternary-Bonsai-4B is the accuracy-per-byte winner above 1 GiB. 83.0% from only 1.1 GiB, within 2 points of Qwen3.5-4B at 40% of the weight size.
they show strong promise on edge devices and where disk space is limited. I think this lab is worth watching.
Nice work, I applied my own benchmarking tools to it.
On my single NVidia Spark I get 173.3 tokens/s on baseline config, 372.4 tokens/s with added tuning/parallel options. Most notably time to first token is incredibly low, similar models take ~6000ms. Bonsai was 70ms (almost 100x reduction) with flash attention
Having said all that, gemma4-e4b-q4km did much better and I can achieve 70% of the tokens/s on the same machine, specifically in context of tool use and for running agents.
This is very interesting and exciting, but IMHO the comparisons read as a bit disingenuous with the other models at 16 bit weights.
Would be interesting to see a comparison to quantized versions of the other models. The 16 bit release of those models is not optimized for size which makes the comparison weaker than it could be.
If this model beats the others also in a fair comparison it gives even more credibility to it.
I think it’s exciting to live in this quirky universe where we have simply accepted our hardware does weird and nonlinear stuff and that powers some math and that’s why your transform function works. Many people thought quantisation is not viable to the extent we see, but we clearly underestimated the effect of hardware on the actual non linearity of the models. Cool to see this pushed to the limits.
Nature has already set an absurdly high bar. The human brain runs on roughly 20 watts, yet delivers a level of intelligence we still can't clearly define, let alone replicate. Nothing we've built comes close... either in capability or efficiency. We're still very early in understanding what "intelligence" even means, much less engineering it. so, we have a long way to go, and push.
Depending on how you convert synapse count to parameters, the brain also has something like a thousand trillion parameters. In that light it's pretty darn surprising that an artificial neural network can produce anything like coherent text.
Maybe the brain is more akin to a network of networks and the actual reasoning part is not all that large? There are lots of areas dedicated exclusively to processing input and controlling subsystems. I can imagine a future where large artificial networks work in a similar way, with multiple smaller ones connected to each other.
It indeed is. We now have models less than 100M params producing pretty coherent, and somewhat relevant text to give input. That is indeed impressive.
I believe the answer lies in how "quickly" (and how?) we are able to learn, and then generalize those learnings as well. As of now, these models need millions (at least) examples to learn, and are still not capable of generalizing the learnings to other domains. Human brains hardly need a few, and then, they generalize those pretty well.
A 1980's desk calculator can multiply two 8 digit numbers with much less energy than your brain takes to do the same.
Modern LLM's similarly beat the human brain in lots of tasks for energy efficiency - mostly by the fact the LLM can produce the answer in 1 second and the brain has to spend half an hour researching and drafting something.
This makes sense. The 1-bit model implies needing 2x as many neurons, because you need an extra level to invert. But the ternary model still has a sign, just really low resolution.
(I've been reading the MMLU-Redux questions for electrical engineering. They're very funny. Fifty years ago they might have been relevant. The references to the Intel 8085 date this to the mid-1970s. Moving coil meters were still a big thing back then. Ward-Leonard drives still drove some elevators and naval guns. This is supposed to be the hand-curated version of the questions. Where do they get this stuff? Old exams?)
So excited to see this - the big advantage of 1.58 bits is there are no multiplications at inference time, so you can run them on radically simpler and cheaper hardware.
At 4 bits, you could just have a hard-wired table lookup. Two 4 bit values in, 256 entry table. You can have saturating arithmetic and a post-processing function for free. Somebody must be building hardware like that.
The raw math: File size is hard-linked to parameter count and quant type. Intelligence is sort of linked to parameter count. Parameter count dictates the hardware requirement. What't left for the labs is, compressing more intelligence into lower parameter count, or packing more of specialized intelligence or buying up more hardware. Those are the only 3 directions all models/labs are heading.
Ever since I saw the first one of these one-bit models made by Microsoft, I thought this was a fascinating route. I assume that in practice, this is less helpful than it seems, just because there's every economic incentive in the world for the big AI labs to produce small, powerful, fast models. None of them seem to be using this technique, so it's interesting, but I suspect it's not quite working.
I also have yet to see any of these at a larger scale. For example, can you try one of these at 100 billion parameters?
The original BitNet was natively trained on 1.58 bits. PrismML has not released any actual info on how they trained, but since they are based on Qwen, there was certainly some downstream quantization involved.
Wouldnt the margin be higher? All other models being moved from unquantized to quantized would lower their performance, while bonsai stays. I get what you see if it was in regards to score/modelsize, but not for absolute performance
The metric they're selling this on is intelligence per byte, rather than total intelligence. So, if they used the quantized competing models, the intelligence per byte gap shrinks, because most models hold up very well down to 6-bit quantization, and 4-bit is usually still pretty good, though intelligence definitely tends to fall below 6-bit.
Nonetheless, the Prism Bonsai models are impressive for their size. Where it falls apart is with knowledge. It has good prose/logic for a tiny model, and it's fast even on modest hardware, but it hallucinates a lot. Which makes sense. You can't fit the world's data in a couple of gigabytes. But, as a base model for fine-tuning for use cases where size matters, it's probably a great choice.
Open access for next 5 hours (Ternary-Bonsai-8B-Q2_0.gguf, running on RTX 3090) or until server crashes or the this spot instance gets taken away :) =>
https://uklkyvetsjf7qt-80.proxy.runpod.net
# llama.cpp is forked one: https://github.com/PrismML-Eng/llama.cpp.git# The server can serve 5 parallel request, with each request capped at around `13K` tokens...
# A bit of of benchmarks I did:
# 1. Input: 1001 tokens, ttfs: 0.3 second, outputs: 1618 tokens ~140t/s
# 2. Input: 9708 tokens, ttfs: 2.4 second, outputs: 2562 tokens at ~106t/s
# Vram usage was consistently at ~7GiB.
> https://huggingface.co/prism-ml/Ternary-Bonsai-8B-gguf/resol...
Fyi, I believe `--flash-attn on` doesn't do anything, you should instead use `--flash-attn 1`. I'm getting ~150t/s on a RTX 3080 10GB as well with f16 cache type.
Thanks.. updated my local docs :)
played around with it a little, works great for - basic conversational tasks - basic reasoning - creating little UI snippets and not good for one shot coding tasks. once it starts hallucinating about something you can't make it rectify that. during a coding task it wrote some madeup library function, despite of pointing it out, giving docs, it kept producing the same code with the madeup function all while acknowledging the problem. btw, thanks for putting it together!
Thanks a lot, I was about to clone their llama.cpp branch and do the same.
Some more interesting tidbits from my go-to tests:
* Fails the car wash test (basic logic seems to be weak in general)
* Fails simple watch face generation in html/css.
* Fails the "how many Rs in raspberry test" (not enough cross-token training data), but will funnily assume you may be talking about Indian Rupees and tell you a lot about raspberry prices in India without being asked. Possible Indian training data unbalance?
* Flat out refuses to talk about Tiananmen square when pushed directly - despite being from a US company. Again, perhaps they are exposed to some censored training data? Anyways, when slowly built up along the conversation by asking about locations and histories, it will eventually tell you about the massacre, so the censorship bias seems weak in general. Also has no problem immediately talking about anything Gaza/Israel/US or other sensitive topics.
* Happily tells you how to synthesize RDX with list of ingredients and chemical process step by step. At least it warns you that it is highly dangerous and legally controlled in the US.
The 1-bit Bonsai and Ternary Bonsai models are all based on the corresponding Qwen3 model: https://raw.githubusercontent.com/PrismML-Eng/Bonsai-demo/re... (page 4)
Thanks, already suspected as much. Also gives context to the other comment here that says it is basically equivalent in accuracy to Qwen3.5-4B. Essentially seems to be a very good quantization of that model, not a new BitNet.
It's a good-per-byte-but-not-in-absolute-terms quantization of Qwen3-8B that's comparable in accuracy to Qwen3.5-4B at 4-bit quantization (which makes the 4B model larger in terms of storage, though the lower number of parameters and hybrid attention give it a speed advantage if you're not bottlenecked on memory bandwidth for the model weights.)
I did a quick benchmark & compared it with Qwen3.5: https://github.com/ArmanJR/PrismML-Bonsai-vs-Qwen3.5-Benchma...
in my results, accuracy-wise Ternary-Bonsai-8B is on par with Qwen3.5-4B. But in accuracy-per-byte, bonsai is the clear winner:
=> Ternary-Bonsai-1.7B achieved 65.1% from 462 MiB, beating Qwen3.5-0.8B by 12 points while being ~5% smaller on disk. => Ternary-Bonsai-4B is the accuracy-per-byte winner above 1 GiB. 83.0% from only 1.1 GiB, within 2 points of Qwen3.5-4B at 40% of the weight size.
they show strong promise on edge devices and where disk space is limited. I think this lab is worth watching.
Nice work, I applied my own benchmarking tools to it.
On my single NVidia Spark I get 173.3 tokens/s on baseline config, 372.4 tokens/s with added tuning/parallel options. Most notably time to first token is incredibly low, similar models take ~6000ms. Bonsai was 70ms (almost 100x reduction) with flash attention
Having said all that, gemma4-e4b-q4km did much better and I can achieve 70% of the tokens/s on the same machine, specifically in context of tool use and for running agents.
Does this sort of thing scale? Would a 30B or higher model see similar performance/memory gains under this scheme?
This is very interesting and exciting, but IMHO the comparisons read as a bit disingenuous with the other models at 16 bit weights. Would be interesting to see a comparison to quantized versions of the other models. The 16 bit release of those models is not optimized for size which makes the comparison weaker than it could be.
If this model beats the others also in a fair comparison it gives even more credibility to it.
I think it’s exciting to live in this quirky universe where we have simply accepted our hardware does weird and nonlinear stuff and that powers some math and that’s why your transform function works. Many people thought quantisation is not viable to the extent we see, but we clearly underestimated the effect of hardware on the actual non linearity of the models. Cool to see this pushed to the limits.
Nature has already set an absurdly high bar. The human brain runs on roughly 20 watts, yet delivers a level of intelligence we still can't clearly define, let alone replicate. Nothing we've built comes close... either in capability or efficiency. We're still very early in understanding what "intelligence" even means, much less engineering it. so, we have a long way to go, and push.
Depending on how you convert synapse count to parameters, the brain also has something like a thousand trillion parameters. In that light it's pretty darn surprising that an artificial neural network can produce anything like coherent text.
Maybe the brain is more akin to a network of networks and the actual reasoning part is not all that large? There are lots of areas dedicated exclusively to processing input and controlling subsystems. I can imagine a future where large artificial networks work in a similar way, with multiple smaller ones connected to each other.
It indeed is. We now have models less than 100M params producing pretty coherent, and somewhat relevant text to give input. That is indeed impressive.
I believe the answer lies in how "quickly" (and how?) we are able to learn, and then generalize those learnings as well. As of now, these models need millions (at least) examples to learn, and are still not capable of generalizing the learnings to other domains. Human brains hardly need a few, and then, they generalize those pretty well.
A 1980's desk calculator can multiply two 8 digit numbers with much less energy than your brain takes to do the same.
Modern LLM's similarly beat the human brain in lots of tasks for energy efficiency - mostly by the fact the LLM can produce the answer in 1 second and the brain has to spend half an hour researching and drafting something.
> Nothing we've built comes close... either in capability or efficiency.
Only when you look at stuff that the brain is specifically good at.
You can surpass the brain with even simple mechanical adders or an abacus in certain subdomains.
General intelligence I mean. What calculations even need to be performed and when, still comes from our brains.
This makes sense. The 1-bit model implies needing 2x as many neurons, because you need an extra level to invert. But the ternary model still has a sign, just really low resolution.
(I've been reading the MMLU-Redux questions for electrical engineering. They're very funny. Fifty years ago they might have been relevant. The references to the Intel 8085 date this to the mid-1970s. Moving coil meters were still a big thing back then. Ward-Leonard drives still drove some elevators and naval guns. This is supposed to be the hand-curated version of the questions. Where do they get this stuff? Old exams?)
[1] https://github.com/aryopg/mmlu-redux/blob/main/outputs/multi...
So excited to see this - the big advantage of 1.58 bits is there are no multiplications at inference time, so you can run them on radically simpler and cheaper hardware.
At 4 bits, you could just have a hard-wired table lookup. Two 4 bit values in, 256 entry table. You can have saturating arithmetic and a post-processing function for free. Somebody must be building hardware like that.
Low level software engineers use look up tables.
Hardware engineers realise that a compiler will almost always find some combination of gates which is smaller/faster than the contents of any table.
A LUT is pretty wasteful. You only have a one bit significand, so the mantissa and sign bits are boolean binops, and the exponent is a 2 bit adder.
and so you can at 1-bit too, and the hardware will be even smaller and cheaper
The raw math: File size is hard-linked to parameter count and quant type. Intelligence is sort of linked to parameter count. Parameter count dictates the hardware requirement. What't left for the labs is, compressing more intelligence into lower parameter count, or packing more of specialized intelligence or buying up more hardware. Those are the only 3 directions all models/labs are heading.
Ever since I saw the first one of these one-bit models made by Microsoft, I thought this was a fascinating route. I assume that in practice, this is less helpful than it seems, just because there's every economic incentive in the world for the big AI labs to produce small, powerful, fast models. None of them seem to be using this technique, so it's interesting, but I suspect it's not quite working.
I also have yet to see any of these at a larger scale. For example, can you try one of these at 100 billion parameters?
How is the research on training these models directly in their quantized state going?
That'll be the real game changer.
The original BitNet was natively trained on 1.58 bits. PrismML has not released any actual info on how they trained, but since they are based on Qwen, there was certainly some downstream quantization involved.
This is pretty cool! I would love to see an even larger models shrunk down.
If you got that into a couple gigs--what could you stuff into 20 gigs?
All of their benchmarks are against 16 bit models right?
Why aren't they comparing to 2/3/4 bit quants?
looked at quant versions of these models and they all outperform it so I guess it just doesn't look as good.
hallucinates in pretty much every answer
Yet again they're comparing against unquantized versions of other models. They would probably still win but by a much smaller size margin.
Wouldnt the margin be higher? All other models being moved from unquantized to quantized would lower their performance, while bonsai stays. I get what you see if it was in regards to score/modelsize, but not for absolute performance
The metric they're selling this on is intelligence per byte, rather than total intelligence. So, if they used the quantized competing models, the intelligence per byte gap shrinks, because most models hold up very well down to 6-bit quantization, and 4-bit is usually still pretty good, though intelligence definitely tends to fall below 6-bit.
Nonetheless, the Prism Bonsai models are impressive for their size. Where it falls apart is with knowledge. It has good prose/logic for a tiny model, and it's fast even on modest hardware, but it hallucinates a lot. Which makes sense. You can't fit the world's data in a couple of gigabytes. But, as a base model for fine-tuning for use cases where size matters, it's probably a great choice.
unfortunately, there doesn't seem to be a clear way to fine-tune these models yet. excited for when that happens though.
installed since last HN post. So Bonsai (1-bit) and Ternary-Bonsai are different?
Can it be run on browsers with WASM/WebGPU?
This model tends to be annoyingly literal. An example from earlier today:
>> What are some names like Llewelyn?
> Some names like Llewelyn are Llewelyn, Llewelyn, Llewelyn, (repeats several times), and Llewelyn.
When do we get 1100B Kimi K2.6 in 160 GB of memory at 1.125 bpw?
> On M4 Pro, Ternary Bonsai 8B runs at 82 toks/sec, roughly 5x faster than a 16-bit 8B model
Wow, if this is true, I am extremely impressed and excited!
I wonder about kv cache how much better it is as well!