I ran it on my laptop, which is a Lenovo Legion 5i (think 32 GB RAM, 4060 w/ 8 GB VRAM, you get the picture). It was a quantized model (otherwise it would not fit on my NVMe 1TB drive) at 4 bits per weight - UD_Q4_K_XL. It ran at about 12 seconds per token (not tokens per second). A fun project, but not worth it. I used 4096 tokens of context cache, and I ran it with llama.cpp - as it supports memory mapping. Because the whole thing could obviously not fit in RAM, I was curious how much it would need to stream from SSD. The answer? For a simple 4 sentence description of who it was, about 1.5 TiB was streamed from disk.
Thank you for sharing. 1.5TB of streamed data at 12 seconds per token on a high end consumer laptop is a pretty high requirement - I can only imagine how much that cost to train. I don't know how running this model could be cost effective for anybody.
Indeed - definitely not cost effective to run it on this laptop LOL. It makes me wonder how fast we could run the model if we could fit the weights entirely within CPU cache (assuming a whole ton of CPUs with low latency & high speed IO of course).
A few people are running highly quantized models with limited context windows. It's still impressive, but not the benchmark level intelligence. Very few people could afford a rig for reasonable local performance at a reasonable quant, at full context size.
The antirez example is 2.6bit quant, 32k context, and few tokens per second... on a ~$7000 MacBook M5 (new RAM pricing).
8 X RTX6000. It will run you around 80-100k to get started with a model at this size with decent tps..
Don't worry though, open source evangelists will tell you that these will be running on your phone in the next 3 years.
For $100k you could run this model 24/7 through open router with 10 concurrent sessions at 50tps for a decade and have money left over for a vacation. There's no point in investing this type of money in local models unless you have a business where you're already paying for many employee's individual token usage.
Not to mention the three separate dedicated 15A circuits you would need to have installed in order to run the 3x 2000W power supplies running ideally at no more than 1400W sustained load each. And definitely would need 200A service to the house if you have a family living there with you.
You can run the NV4FP quant with 8x RTX6000 cards at 50-75 tps output, but not (practically speaking) the OEM FP8 version. You will learn more about PCIe than you ever wanted to know.
The real gangstas are running 16x RTX6000s. Too rich for my blood, and the NV4FP quant doesn't seem to be that much worse.
Depends how much you value privacy and running uncensored models.
Personally, I’m waiting for hardware to hit the secondary market before I buy something to run unquantized models like GLM. But I have no doubt that I will, at some point.
>Don't worry though, open source evangelists will tell you that these will be running on your phone in the next 3 years.
Not sure if you're being sarcastic, but I can run a quantised version of Gemma or Qwen on my 16GB M1 Macbook Pro that beats GPT-4 from 2023 hands-down.
I wouldn't be surprised if, in another 3 years, you'd be able to run something as powerful as Opus 4.5 or GLM-5.2 on standard consumer hardware - say a 32GB/64GB M7 Pro.
I also wouldn't be surprised if, 3 years after that, cheaper hardware and improved model efficiency means that there's a much smaller gap between what you can run on a consumer CPU (which, with memory prices coming down, could look like a 256GB M9 or M10 Pro) and $100k GPU cluster.
This is clearly where the industry is going, imho. Everyone who is playing with LLMs wants a laptop with enough grunt to run a decent model locally.
We've been sat with basically the same PC specs for ~20 years - our current specs are within an order of magnitude of the ones we could buy back in 2010. This is not really constrained by tech, as we could have much, much, larger machines. It's more because there's no mass demand for much, much, larger machines - if it's big enough to run Office apps or VSCode then you're good to go. The exponential growth we saw in the 90's was driven as much by software demand as it was by hardware development.
I can see the next 10 years produce the same kind of push for larger machines that the 90's did. And we should probably expect the same kind of standards churn as our existing technologies for storage, memory, etc, don't scale up enough and new technologies become worth developing because there's demand for them.
I suspect AI labs are buying stuff not just for their own use, but to make local use too expensive to be an option :-( And they can always make the "best" frontier model even bigger (though only fractionally better) so it's always out of reach of local use, while consumer laptops have nearly the same amount of memory they had a decade ago.
m o
o
d
e
l o
s
i o
z o
e 2020 2022 2024 2026
c
h
e
a
p o
R o
A o
M o
2020 2022 2024 2026
I'm being very sarcastic, local model evangalists seems to just be operating on vibes when they say these things and are completely disconnected from how models work, what the hardware requirements are.
Prices aren't going down, and consumer platforms are being shipped with less RAM so we can be sold cloud products. This isn't going to happen.
Can you please explain to me how you're going to fit 700bb-1T params in 64GB of RAM? You realize there are memory requirements proportional to model size?
Would you be better off pooling that money with some hackerspace group and then setting up shared inference infra, so that way you at least get better utilization?
How do the economics of your statement work out? Clearly inference providers don't have a time to ROI of 10 years on their hardware costs; and that's without even taking ongoing energy costs into account. What's missing here?
Inference providers have been getting a firehose of investor cash to keep the chips running (and are looking around very nervously as that firehose starts to sputter).
Yeah, the neoclouds and hyperscalers are taking massive losses right now, self hosting is basically signing yourself up to do the same. There are philosophical reasons to do so but it’s a terrible economic decision
Yea as far has hobbies go, I feel like this is on the low end. I know people who collect watches and corvettes, that's way more expensive and functionally you can't really do anything special with them.
Sure, If you want to light money on fire for entertainment, more power to you. There's probably worse ways to light 100k on fire. If I have an extra 100k laying around it's going to my family though.
Yeah I completely agree. But this is much larger model than the 8B one they put on a chip, so that's probably an engineering challenge for now. Also, how expensive would it be?
As an individual I do not need the whole model. I don't need the model to have knowledge of the rain history of Algeria nor how many colors are in the Russian flag. Once they start trimming down the excess and making them field focused they will run just fine on people's individual devices.
> I do not need the whole model. I don't need the model to have knowledge of the rain history of Algeria nor how many colors are in the Russian flag
Isn’t the performance gap between quantized and full models indicative that even if you aren’t using it directly, the model knowing the colors in the Russian flag does have something to do with the intelligence you demand?
I ran it on my laptop, which is a Lenovo Legion 5i (think 32 GB RAM, 4060 w/ 8 GB VRAM, you get the picture). It was a quantized model (otherwise it would not fit on my NVMe 1TB drive) at 4 bits per weight - UD_Q4_K_XL. It ran at about 12 seconds per token (not tokens per second). A fun project, but not worth it. I used 4096 tokens of context cache, and I ran it with llama.cpp - as it supports memory mapping. Because the whole thing could obviously not fit in RAM, I was curious how much it would need to stream from SSD. The answer? For a simple 4 sentence description of who it was, about 1.5 TiB was streamed from disk.
Thank you for sharing. 1.5TB of streamed data at 12 seconds per token on a high end consumer laptop is a pretty high requirement - I can only imagine how much that cost to train. I don't know how running this model could be cost effective for anybody.
Indeed - definitely not cost effective to run it on this laptop LOL. It makes me wonder how fast we could run the model if we could fit the weights entirely within CPU cache (assuming a whole ton of CPUs with low latency & high speed IO of course).
short answer: they mostly aren't
A few people are running highly quantized models with limited context windows. It's still impressive, but not the benchmark level intelligence. Very few people could afford a rig for reasonable local performance at a reasonable quant, at full context size.
The antirez example is 2.6bit quant, 32k context, and few tokens per second... on a ~$7000 MacBook M5 (new RAM pricing).
8 X RTX6000. It will run you around 80-100k to get started with a model at this size with decent tps..
Don't worry though, open source evangelists will tell you that these will be running on your phone in the next 3 years.
For $100k you could run this model 24/7 through open router with 10 concurrent sessions at 50tps for a decade and have money left over for a vacation. There's no point in investing this type of money in local models unless you have a business where you're already paying for many employee's individual token usage.
> 8 X RTX6000. It will run you around 80-100k to get started
8 x RTX6000 GPUs cost $100,000 alone. You then need to build a system that can support those GPUs with enough PCIe lanes through a PCIe switch.
It's going to be $120K to $150K to build or buy a system to run this.
Not to mention the three separate dedicated 15A circuits you would need to have installed in order to run the 3x 2000W power supplies running ideally at no more than 1400W sustained load each. And definitely would need 200A service to the house if you have a family living there with you.
But hey you could save on heating?
4 replies →
You can run the NV4FP quant with 8x RTX6000 cards at 50-75 tps output, but not (practically speaking) the OEM FP8 version. You will learn more about PCIe than you ever wanted to know.
The real gangstas are running 16x RTX6000s. Too rich for my blood, and the NV4FP quant doesn't seem to be that much worse.
7 replies →
isn't throwing that into a [insert financial vehicle that gives 99.99999% safe returns] going to destroy that when you factor in electricity costs?
Or even just electricity costs vs token cost
Depends how much you value privacy and running uncensored models.
Personally, I’m waiting for hardware to hit the secondary market before I buy something to run unquantized models like GLM. But I have no doubt that I will, at some point.
>Don't worry though, open source evangelists will tell you that these will be running on your phone in the next 3 years.
Not sure if you're being sarcastic, but I can run a quantised version of Gemma or Qwen on my 16GB M1 Macbook Pro that beats GPT-4 from 2023 hands-down.
I wouldn't be surprised if, in another 3 years, you'd be able to run something as powerful as Opus 4.5 or GLM-5.2 on standard consumer hardware - say a 32GB/64GB M7 Pro.
I also wouldn't be surprised if, 3 years after that, cheaper hardware and improved model efficiency means that there's a much smaller gap between what you can run on a consumer CPU (which, with memory prices coming down, could look like a 256GB M9 or M10 Pro) and $100k GPU cluster.
This is clearly where the industry is going, imho. Everyone who is playing with LLMs wants a laptop with enough grunt to run a decent model locally.
We've been sat with basically the same PC specs for ~20 years - our current specs are within an order of magnitude of the ones we could buy back in 2010. This is not really constrained by tech, as we could have much, much, larger machines. It's more because there's no mass demand for much, much, larger machines - if it's big enough to run Office apps or VSCode then you're good to go. The exponential growth we saw in the 90's was driven as much by software demand as it was by hardware development.
I can see the next 10 years produce the same kind of push for larger machines that the 90's did. And we should probably expect the same kind of standards churn as our existing technologies for storage, memory, etc, don't scale up enough and new technologies become worth developing because there's demand for them.
4 replies →
> memory prices coming down
Are they?
I suspect AI labs are buying stuff not just for their own use, but to make local use too expensive to be an option :-( And they can always make the "best" frontier model even bigger (though only fractionally better) so it's always out of reach of local use, while consumer laptops have nearly the same amount of memory they had a decade ago.
For most tasks, I don't value the LLMs based on their absolute capabilities. I wouldn't want to use GPT-4 today even if it's free.
I'm being very sarcastic, local model evangalists seems to just be operating on vibes when they say these things and are completely disconnected from how models work, what the hardware requirements are.
Prices aren't going down, and consumer platforms are being shipped with less RAM so we can be sold cloud products. This isn't going to happen.
Can you please explain to me how you're going to fit 700bb-1T params in 64GB of RAM? You realize there are memory requirements proportional to model size?
1 reply →
Would you be better off pooling that money with some hackerspace group and then setting up shared inference infra, so that way you at least get better utilization?
And before you know it, you invented some openrouter provider from first principles...
1 reply →
You can then rent spare capacity out to people on a subscription or token basis ….wait
How do the economics of your statement work out? Clearly inference providers don't have a time to ROI of 10 years on their hardware costs; and that's without even taking ongoing energy costs into account. What's missing here?
Output tokens are actually kinda expensive for the provider.
The input cache hit tokens are incredibly cheap for them, (incredibly high margin too, except for deepseek).
And input tokens are in the middle. Input tokens can be processed very efficiently.
Also his math is wrong. $100k gets you 22.7B output tokens at $4.4/M which is how much GLM 5.2 costs.
At 500/s 22.7B is just 500 days. Or about 1.54 years. Which is much less then the life of the hardware.
The inference providers are running batch sizes much larger than 10
Inference providers have been getting a firehose of investor cash to keep the chips running (and are looking around very nervously as that firehose starts to sputter).
https://aimultiple.com/gpu-benchmark
concurrency
Yeah, the neoclouds and hyperscalers are taking massive losses right now, self hosting is basically signing yourself up to do the same. There are philosophical reasons to do so but it’s a terrible economic decision
you can however, have fun with it.
oil workers buy 100k trucks they do not-much with. why not a 100k in computer?
I can't help but ask where this comment came from, you must have some exposure..
5 replies →
Yea as far has hobbies go, I feel like this is on the low end. I know people who collect watches and corvettes, that's way more expensive and functionally you can't really do anything special with them.
6 replies →
Because car loans can’t be used to buy computers
4 replies →
Sure, If you want to light money on fire for entertainment, more power to you. There's probably worse ways to light 100k on fire. If I have an extra 100k laying around it's going to my family though.
Given GLM is open weight - all you need is one company to take the taalas approach ( model on hardware ), and you're sorted right?
https://taalas.com/products/
Yeah I completely agree. But this is much larger model than the 8B one they put on a chip, so that's probably an engineering challenge for now. Also, how expensive would it be?
1 reply →
As an individual I do not need the whole model. I don't need the model to have knowledge of the rain history of Algeria nor how many colors are in the Russian flag. Once they start trimming down the excess and making them field focused they will run just fine on people's individual devices.
> I do not need the whole model. I don't need the model to have knowledge of the rain history of Algeria nor how many colors are in the Russian flag
Isn’t the performance gap between quantized and full models indicative that even if you aren’t using it directly, the model knowing the colors in the Russian flag does have something to do with the intelligence you demand?
13 replies →
> 50tps for a decade
assuming demand doesn't keep on increasing. even google has trouble having enough capacity apparently.
Or you have data that HIPAA, GDPR, PII, or have to care about the concern of others training on your data.
That too.
[dead]
Run quantized versions. https://unsloth.ai/docs/models/glm-5.2
follow antirez - https://x.com/antirez/status/2071173841175363905?s=20
https://xcancel.com/antirez/status/2071173841175363905
Thats quantized
It's a nice technical achievement but looks unusably slow for actual work