The hardware situation is way better than you think, and quantization is a huge part of why.
Take Qwen 3.5 27B, which is a solid coding model. At FP16 it needs 54GB of VRAM. Nobody's running that on consumer hardware. At Q4_K_M quantization, it needs 16GB. A used RTX 3090 has 24GB and goes for about $900. That model runs locally with room for context.
For 14B coding models at Q4, you're looking at about 10GB. A used RTX 3060 12GB handles that for under $270.
The gap between "needs a datacenter" and "runs on my desk" is almost entirely quantization. A 27B model at Q4 loses surprisingly little quality for most coding tasks. It's not free, but it's not an RTX 7090 either. A used 3090 is probably the most recommended card in the local LLM community right now, and for good reason.
14B even at Q4 isn't realistic for coding on a single 12GB RTX 3060. Token speed is too slow. After all they are dense models. You aren't getting a good MoE model under 30B. You can do OCR, STT, TTS really well and for LLMs, good use cases are classification, summarization and extraction with <10B models.
Remaining dependent on proprietary frontier models that you can only access via an API makes no sense whatsoever. My hope is that the future is open weight models running on local hardware.
"This is what's happening to the parameters of models when they're quantized down to sizes that are possible to run on your laptop. Instead of floats, small integers are what get stored and loaded into memory. When the time comes to use the quantized values, to generate an answer to a question for example, the values are dequantized on the fly. You might think this sounds slower, but we'll see later on that this actually ends up being faster as well as smaller."
I thought that most GPUs supported floating point math in these quantized formats, like they can natively do math on an float4 number (that's maybe packed, 2 float4s into a single byte, or more probably 16 float4s in an 8 byte array or maybe something even bigger)
Am I getting this wrong - is it instead the GPU pulls in the quantized numbers and then converts them back into 32-bit or 64-bit float to actually run through the ALUs on the GPU? (and the memory bandwidth savings make up for the extra work to convert them back into 32 bit numbers once you get them onto the GPU?)
Or is it some weird hybrid, like there is native support for float8 and Bfloat16, but if you want to use float2 you have to convert it to float4 or something the hardware can work with.
I am confused what actually happens in the vectorized ADD and MULT instructions in the GPU with these quantized numbers.
> I am confused what actually happens in the vectorized ADD and MULT instructions in the GPU with these quantized numbers.
I might be wrong, but I think LLM is all about comparing distance between tokens. You can tell that -255 and +255 are very separated, but you are also away that -8 and +8 are also very far away.
Microsoft Bitnet and Google TurboQuant shows that in extreme you can use just -1, 0, +1
Very old CPUs had support only down to FP16, which is useful in graphics applications.
Then support for Bfloat16 and for INT8 has been added, which are not useful for anything else but AI/ML applications. Then support for FP8 has been added. Even smaller formats are supported only on some very recent GPUs.
If you have a recent enough GPU, it might support something like float2 or float4, but if you have an older GPU you must convert the short format to the next bigger format that is supported, before performing some operations.
Hardware support will vary widely, as will speed on these smaller FP formats, sometimes intentionally nerfed in consumer cards.
Lots of devices with embedded "AI accelerators" will also only do things like INT8, and for some reason INT8 is generally worse than the same size FP8 (maybe that could be fixed with smarter quantization).
Leading to my question: Ok keeping a zero and a minus-zero does make sense for some limits calculations...
But when all you have is 4 bits, is this not quite wasteful? Would using the bits for eg. a 2.5 not improve the model?
It might be useful. The Lion optimizer uses 1-bit values to represent forward or backward. NNs can pick up on patterns like that in very strange ways. Of course, those are 1's, not 0's, so maybe the benefit disappears when multiplying by zero. But it's important to challenge assumptions like "well, let's get rid of the negative half of 0" before you test experimentally whether it's useful or not. NNs are nothing if not shockingly weird when you try to make them.
This is beautifully written and visualised, well done! The KL divergence comparisons between original and different quantisation levels is on-point. I'm not sure people realize how powerful quantisation methods are and what they've done for democratising local AI. And there are some great players out there like Unsloth and Pruna.
Thank you! I was really surprised how robust models are to losing information. It seems wrong that they can be compressed so much and still function at all, never mind function quite closely to the original size.
Think we're only going to keep seeing more progress in this area on the research side, too.
In my personal opinion I don’t think the 1.58 bit work is going to make it into the mainstream.
Not sure why you think fractional representations are only useful for training? Being able to natively compute in lower precisions can be a huge performance boost at inference time.
I read the entire thing top-to-bottom, as a visual learner this is superb.
One nitpick -- in the "asymmetric quantification" code, shouldn't "zero" be called "midpoint" or similar? Or is "zero" an accepted mathematics term in this domain?
“Zero point” is how I saw it referred to in the literature, so that’s what I went with. I personally prefer to think of it as an offset, but I try to stick with terms folks are likely to see in the wild.
Quantization is important for me because it's the only way out I can see for a future of programming that doesn't involve going through a giant bigco who can run, as the article says, a machine with 2TB of memory. And not just memory, but my understanding is that for the model to be performant, it has to be VRAM to boot.
This comes as the latest concern of mine in a long line around "how software gets written" remaining free-as-in-freedom. I've always been really uneasy about how reliant many programming languages were on Jetbrains editors, only vaguely comforted by their "open-core" offering, which naturally only existed for languages with strong OSS competition for IDEs (so... java and python, really). "Intellisense" seemed very expensive to implement and was hugely helpful in writing programs without stopping every 4 seconds to look up whether removing whitespace at the end of a line is trim, strip, or something else in this language. I was naturally pleased to see language servers take off, even if it was much to my chagrin that it came from Microsoft, who clearly was out of open standards to EEE and decided to speed up the process by making some new ones.
Now LLMs are the next big worry of mine. It seems pretty bad for free and open software if the "2-person project, funded indirectly by the welfare state of a nordic or eastern-european nation" model that drives ridiculously important core libre/OSS libraries now is even less able to compete with trillion dollar corporations.
Open-weight, quantized, but still __good__ models seem like the only way out. I remain somewhat hopeful just from how far local models have come - they're significantly more usable than they were a year ago, and we've got more tools like LM Studio etc making running them easy. But there's still a good way to go.
I'll be sad if a "programming laptop" ends up going from "literally anything that can run debian" to "yeah you need an RTX 7090, 128GB of VRAM, and the 2kW wearable power supply backpack addon at a minimum".
Sam's previous posts are well worth digging up too. This one is outstanding, but they're all good. I really enjoyed this and learned a lot.
I'm a bit envious of his job. Learning to teach others, and building out such cool interactive, visual documents to do it? He makes it look easier than it is, of course. A lot of effort and imagination went into this, and I'm sure it wasn't a walk in the park. Still, it seems so gratifying.
The 2 bit is probably slower because it clashes with some register sizes and how data is read in blocks. No additional benefit because the architecture doesn't read 2 bits but probably min 4 bits and then it clashes with utilization.
One thing from practical experience - the quality gap between model sizes shows up in a way benchmarks don't capture. I have a system where a smaller model generates plans and a larger model can override them. On any single output they look comparable. The difference shows up 3-4 steps later — small model makes a decision that sounds reasonable but compounds into a bad plan. Perplexity won't catch that, KL divergence won't either. They both measure one prediction at a time.
What is the best way to archive a JS heavy site like this? I reviewed OPs github and they haven't open-sourced these visualizations probably because they are tied to his employer.
Most (all?) of this holds for quantizing convnets too, if you're looking for an easy exercise you can play around with quantizing resnet50 or something and plotting layer activations
something I have been wondering about is doing regressive layer specific quantization based on large test sets. ie reduce very specifically layers that don't improve general quality.
59 comments
Take Qwen 3.5 27B, which is a solid coding model. At FP16 it needs 54GB of VRAM. Nobody's running that on consumer hardware. At Q4_K_M quantization, it needs 16GB. A used RTX 3090 has 24GB and goes for about $900. That model runs locally with room for context.
For 14B coding models at Q4, you're looking at about 10GB. A used RTX 3060 12GB handles that for under $270.
The gap between "needs a datacenter" and "runs on my desk" is almost entirely quantization. A 27B model at Q4 loses surprisingly little quality for most coding tasks. It's not free, but it's not an RTX 7090 either. A used 3090 is probably the most recommended card in the local LLM community right now, and for good reason.
Add a third one and you can run Qwen 3.5 27B Q6 with 128k ctx. For less than the price of a 3090.
> 3x RTX 3060 less tgab the price of a 3090
Interesting, here it is around the same. 200-250€ for a used 12GB 3060 and 600-800 for a used 3090€.
Running LLM makes no sense whatsoever
https://github.com/z-lab/paroquant
"This is what's happening to the parameters of models when they're quantized down to sizes that are possible to run on your laptop. Instead of floats, small integers are what get stored and loaded into memory. When the time comes to use the quantized values, to generate an answer to a question for example, the values are dequantized on the fly. You might think this sounds slower, but we'll see later on that this actually ends up being faster as well as smaller."
I thought that most GPUs supported floating point math in these quantized formats, like they can natively do math on an float4 number (that's maybe packed, 2 float4s into a single byte, or more probably 16 float4s in an 8 byte array or maybe something even bigger)
Am I getting this wrong - is it instead the GPU pulls in the quantized numbers and then converts them back into 32-bit or 64-bit float to actually run through the ALUs on the GPU? (and the memory bandwidth savings make up for the extra work to convert them back into 32 bit numbers once you get them onto the GPU?)
Or is it some weird hybrid, like there is native support for float8 and Bfloat16, but if you want to use float2 you have to convert it to float4 or something the hardware can work with.
I am confused what actually happens in the vectorized ADD and MULT instructions in the GPU with these quantized numbers.
M1 Max: FP16 hardware support, FP8 and Bfloat16 emulated in software (via dequantization)
H100: FP16 and FP8 hardware support
> which I ran both on a MacBook Pro M1 Max and a rented H100 SXM GPU
> I am confused what actually happens in the vectorized ADD and MULT instructions in the GPU with these quantized numbers.
I might be wrong, but I think LLM is all about comparing distance between tokens. You can tell that -255 and +255 are very separated, but you are also away that -8 and +8 are also very far away.
Microsoft Bitnet and Google TurboQuant shows that in extreme you can use just -1, 0, +1
Then support for Bfloat16 and for INT8 has been added, which are not useful for anything else but AI/ML applications. Then support for FP8 has been added. Even smaller formats are supported only on some very recent GPUs.
If you have a recent enough GPU, it might support something like float2 or float4, but if you have an older GPU you must convert the short format to the next bigger format that is supported, before performing some operations.
Lots of devices with embedded "AI accelerators" will also only do things like INT8, and for some reason INT8 is generally worse than the same size FP8 (maybe that could be fixed with smarter quantization).
Think we're only going to keep seeing more progress in this area on the research side, too.
To Nemotron 3 Super, which had 25T of nvfp4 native pretraining! https://docs.nvidia.com/nemotron/0.1.0/nemotron/super3/pretr...
Hopefully Microsoft keeps pushing BitNet too, so only "1.58" bits are needed.
I think fractional representations are only relevant for training at this point, and bf16 is sufficient, no need for fp4 and such.
In my personal opinion I don’t think the 1.58 bit work is going to make it into the mainstream.
Not sure why you think fractional representations are only useful for training? Being able to natively compute in lower precisions can be a huge performance boost at inference time.
> Learned rotations for INT4 are cool! Seems similar to SpinQuant?
https://arxiv.org/abs/2405.16406Indeed, but much better! More accurate, less time and space overhead, beats AWQ on almost every bench. I hope it becomes the standard.
> In my personal opinion I don’t think the 1.58 bit work is going to make it into the mainstream.
I hope you're wrong! I'm more optimistic. Definitely a bit more work to be done, but still very promising.
> Being able to natively compute in lower precisions can be a huge performance boost at inference time.
ParoQuant is barely worse than FP16. Any less precise fractional representation is going to be worse than just using that IMO.
One nitpick -- in the "asymmetric quantification" code, shouldn't "zero" be called "midpoint" or similar? Or is "zero" an accepted mathematics term in this domain?
This comes as the latest concern of mine in a long line around "how software gets written" remaining free-as-in-freedom. I've always been really uneasy about how reliant many programming languages were on Jetbrains editors, only vaguely comforted by their "open-core" offering, which naturally only existed for languages with strong OSS competition for IDEs (so... java and python, really). "Intellisense" seemed very expensive to implement and was hugely helpful in writing programs without stopping every 4 seconds to look up whether removing whitespace at the end of a line is trim, strip, or something else in this language. I was naturally pleased to see language servers take off, even if it was much to my chagrin that it came from Microsoft, who clearly was out of open standards to EEE and decided to speed up the process by making some new ones.
Now LLMs are the next big worry of mine. It seems pretty bad for free and open software if the "2-person project, funded indirectly by the welfare state of a nordic or eastern-european nation" model that drives ridiculously important core libre/OSS libraries now is even less able to compete with trillion dollar corporations.
Open-weight, quantized, but still __good__ models seem like the only way out. I remain somewhat hopeful just from how far local models have come - they're significantly more usable than they were a year ago, and we've got more tools like LM Studio etc making running them easy. But there's still a good way to go.
I'll be sad if a "programming laptop" ends up going from "literally anything that can run debian" to "yeah you need an RTX 7090, 128GB of VRAM, and the 2kW wearable power supply backpack addon at a minimum".
I'm a bit envious of his job. Learning to teach others, and building out such cool interactive, visual documents to do it? He makes it look easier than it is, of course. A lot of effort and imagination went into this, and I'm sure it wasn't a walk in the park. Still, it seems so gratifying.
Really good visualizations overall.
One thing from practical experience - the quality gap between model sizes shows up in a way benchmarks don't capture. I have a system where a smaller model generates plans and a larger model can override them. On any single output they look comparable. The difference shows up 3-4 steps later — small model makes a decision that sounds reasonable but compounds into a bad plan. Perplexity won't catch that, KL divergence won't either. They both measure one prediction at a time.