You seem to be unaware that it only takes about four NVIDIA HGX H100 nodes (32 GPUs) to train something like qwen3.5:122b. That model is about as good as ChatGPT was six months to a year ago (for the usual use cases). That would take a long ass time though (over a year) so you’d want probably 50-100 HGX H100s (or lots of the newer, cheaper ARM-based hardware devices).
The weights for qwen3.5:122b are open. That means that if you’ve got the hardware (loads of universities and non-profits have waaaay TF more than 4 HGX H100 nodes) you can continue modern AI development. Everything you need is right there on Huggingface! Deepseek’s stuff is also open I think but I forget. Aside: In my head, I hold the qwen models as “the gold standard” based on many articles I’ve read about them but AI moves so fast, there might be better stuff out on any given day! I haven’t read AI news in like a week so I could be all wrong and qwen3.5 is now sooo obsolete, hehe (that’s how it feels to follow AI news, anyway 🤣).
Even more interesting: qwen3.5:122b isn’t just an LLM. It does visual reasoning (e.g. give it a picture of a plant and ask it to identify it, count the number of screws in an image, estimate distances, etc) as well as the usual LLM stuff. You can read all about it here:
…and if you install ollama and spend $20 on ollama.com’s cloud service you can actually try it out without having to own enough GPUs to cover the 245+GB requirement. I highly recommend that service! You can try out all the latest & greatest models on your local PC (or phone!) for any purpose you want for a $20. Whenever a new model is out they usually have it up on their servers within a day or two and it’s fast, too.
FYI: I’ve used ollama cloud to evaluate models for coding (web dev with Python back end) and qwen3.5:122b is fantastic. It’s not as good as Claude Opus 4.6 but it’s close (and cheap) enough that you can just make up for the mistakes with extra instances that check the output with a critical eye (the latest trick in AI-based coding to get good output).
For reference, the University of Texas at Austin has data centers with 4,000 NVIDIA Blackwell (B200/GB200) GPUs, Harvard has 1,144 GPUs, and the University of Cambridge & Bristol (in the UK) has some monstrous mix of Intel and AMD GPUs. All three are perfectly capable of training new models from scratch or using continuing development on existing open-weight models like Deepseek and Qwen.
Generative AI isn’t going anywhere. Furthermore, advancements in that space happen so fast that it’s likely that in a few years we won’t need so many GPUs/VRAM to train models. Especially if ternary models (and similar, like Google’s TurboQuant tech) take off.
I know this is a long comment but I want to point something else out: If OpenAI and Anthropic go bust, that would flood the market with cheap GPUs. It would be a total price collapse and you can bet your ass that clever universities and service providers (like Amazon compute, but 3rd party) would snap those up and bring down prices across the board.
I’m sorry if I was unclear the first two times I asked, but when I said:
care to link me to all these great models from academics and open-source institutions?
I was interested in the models you’re currently using, not the ones you’re speculating about. Hopefully it goes without saying that “open” weights are precompiled closed-source blobs, and "Open"AI is anything but, etc.
I’m aware new models are trained at the speed of light and hardware is going obsolete faster than it can be put on racks, which is already a problem, so I would love to believe your theory about inexpensive AI GPUs but those very same companies are already going into debt without selling their current stock.
This is not speculation. That’s what I’m actually using nearly every day. It’s not as good as Claude Code with Opus 4.6 but it’s about 90% of the way there (if you use it right). When GLM-5 came out that’s when I cancelled my Claude subscription and just stuck with Ollama Cloud.
I can use gpt-oss:20b on my GPU (4060 Ti 16GB)—and it works well—but for $20/month, the ability to use qwen3.5 and GLM-5 are better options.
I still use my GPU for (serious) image generation though. Using ChatGPT (DALL-E) or Gemini (Nano Banana) are OK for one-offs but they’re slow AF compared to FLUX 2 and qwen’s image models running locally. I can give it a prompt and generate 32 images in no time, pick the best one, then iterate from there (using some sophisticated ComfyUI setups). The end result is a superior image than what you’d get from Big AI.
I’ve used them both for coding and they work really well (way better than you’d think). They’re also perfectly capable of the usual LLM chat stuff (e.g. check my grammar) but all the models (even older, smaller ones) are capable of that stuff these days.
For a treat: Have someone show you using some of these models to search the web! It’s amazing. You don’t see ads, you don’t have to comb through 12 pages of search results, and they read the pages that moment (not cached) to give you summaries of the content. So when you click the link to go to the content you know it’s the thing you were looking for. They’re not using a local index of the Internet, they’re searching on your behalf using whatever search engines you configured. It’s waaaaay better than ChatGPT (which uses Bing behind the scenes whether you like it or not) or Gemini (which uses Google, obviously). The (self-hosted) LLM will literally be running curl for you on Google, DuckDuckGo, Bing, or whatever TF else you want (simultaneously) then reading each of the search results and using your prompt to figure out what the most relevant results are. It’s sooooo nice!
FYI: Ollama.com’s library page is actually a great resource for finding info on all the models that can be self-hosted: https://ollama.com/library
What inclined you to @ me into this? As far as I can see, I haven’t even replied in this thread, and you just seem like you’re on the warpath with anyone that wants to defend using LLMs. If Greg KH thinks it’s coming into it’s own, you might want to heed him.
Before we spin into a tangent about theory and “what ifs” etc, care to link me to all these great models from academics and open-source institutions?
Because right now, the only companies I see making advancements in “AI” are burning through obscene amounts of cash, with no end in sight.
And there is no evidence the cost of inference is going down, and even Anthropic admits training will continue burning resources.
You seem to be unaware that it only takes about four NVIDIA HGX H100 nodes (32 GPUs) to train something like qwen3.5:122b. That model is about as good as ChatGPT was six months to a year ago (for the usual use cases). That would take a long ass time though (over a year) so you’d want probably 50-100 HGX H100s (or lots of the newer, cheaper ARM-based hardware devices).
The weights for qwen3.5:122b are open. That means that if you’ve got the hardware (loads of universities and non-profits have waaaay TF more than 4 HGX H100 nodes) you can continue modern AI development. Everything you need is right there on Huggingface! Deepseek’s stuff is also open I think but I forget. Aside: In my head, I hold the qwen models as “the gold standard” based on many articles I’ve read about them but AI moves so fast, there might be better stuff out on any given day! I haven’t read AI news in like a week so I could be all wrong and qwen3.5 is now sooo obsolete, hehe (that’s how it feels to follow AI news, anyway 🤣).
Even more interesting: qwen3.5:122b isn’t just an LLM. It does visual reasoning (e.g. give it a picture of a plant and ask it to identify it, count the number of screws in an image, estimate distances, etc) as well as the usual LLM stuff. You can read all about it here:
https://ollama.com/library/qwen3.5:122b
…and if you install
ollamaand spend $20 on ollama.com’s cloud service you can actually try it out without having to own enough GPUs to cover the 245+GB requirement. I highly recommend that service! You can try out all the latest & greatest models on your local PC (or phone!) for any purpose you want for a $20. Whenever a new model is out they usually have it up on their servers within a day or two and it’s fast, too.FYI: I’ve used ollama cloud to evaluate models for coding (web dev with Python back end) and qwen3.5:122b is fantastic. It’s not as good as Claude Opus 4.6 but it’s close (and cheap) enough that you can just make up for the mistakes with extra instances that check the output with a critical eye (the latest trick in AI-based coding to get good output).
For reference, the University of Texas at Austin has data centers with 4,000 NVIDIA Blackwell (B200/GB200) GPUs, Harvard has 1,144 GPUs, and the University of Cambridge & Bristol (in the UK) has some monstrous mix of Intel and AMD GPUs. All three are perfectly capable of training new models from scratch or using continuing development on existing open-weight models like Deepseek and Qwen.
Generative AI isn’t going anywhere. Furthermore, advancements in that space happen so fast that it’s likely that in a few years we won’t need so many GPUs/VRAM to train models. Especially if ternary models (and similar, like Google’s TurboQuant tech) take off.
I know this is a long comment but I want to point something else out: If OpenAI and Anthropic go bust, that would flood the market with cheap GPUs. It would be a total price collapse and you can bet your ass that clever universities and service providers (like Amazon compute, but 3rd party) would snap those up and bring down prices across the board.
I’m sorry if I was unclear the first two times I asked, but when I said:
I was interested in the models you’re currently using, not the ones you’re speculating about. Hopefully it goes without saying that “open” weights are precompiled closed-source blobs, and "Open"AI is anything but, etc.
I’m aware new models are trained at the speed of light and hardware is going obsolete faster than it can be put on racks, which is already a problem, so I would love to believe your theory about inexpensive AI GPUs but those very same companies are already going into debt without selling their current stock.
I literally said I’m using qwen3.5:122b for coding. I also use GLM-5 but it’s slightly slower so I generally stick with qwen.
It’s right there, in ollama’s library: https://ollama.com/library/qwen3.5:122b
The weights and everything else for it are on Huggingface: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
This is not speculation. That’s what I’m actually using nearly every day. It’s not as good as Claude Code with Opus 4.6 but it’s about 90% of the way there (if you use it right). When GLM-5 came out that’s when I cancelled my Claude subscription and just stuck with Ollama Cloud.
I can use gpt-oss:20b on my GPU (4060 Ti 16GB)—and it works well—but for $20/month, the ability to use qwen3.5 and GLM-5 are better options.
I still use my GPU for (serious) image generation though. Using ChatGPT (DALL-E) or Gemini (Nano Banana) are OK for one-offs but they’re slow AF compared to FLUX 2 and qwen’s image models running locally. I can give it a prompt and generate 32 images in no time, pick the best one, then iterate from there (using some sophisticated ComfyUI setups). The end result is a superior image than what you’d get from Big AI.
@ikidd@lemmy.world @ ingeanus@ttrpg.network do you two have a source for these supposed great models?
I personally love glm-5 and qwen3.5, specifically: https://ollama.com/library/qwen3.5:122b
I’ve used them both for coding and they work really well (way better than you’d think). They’re also perfectly capable of the usual LLM chat stuff (e.g. check my grammar) but all the models (even older, smaller ones) are capable of that stuff these days.
For a treat: Have someone show you using some of these models to search the web! It’s amazing. You don’t see ads, you don’t have to comb through 12 pages of search results, and they read the pages that moment (not cached) to give you summaries of the content. So when you click the link to go to the content you know it’s the thing you were looking for. They’re not using a local index of the Internet, they’re searching on your behalf using whatever search engines you configured. It’s waaaaay better than ChatGPT (which uses Bing behind the scenes whether you like it or not) or Gemini (which uses Google, obviously). The (self-hosted) LLM will literally be running
curlfor you on Google, DuckDuckGo, Bing, or whatever TF else you want (simultaneously) then reading each of the search results and using your prompt to figure out what the most relevant results are. It’s sooooo nice!FYI: Ollama.com’s library page is actually a great resource for finding info on all the models that can be self-hosted: https://ollama.com/library
What inclined you to @ me into this? As far as I can see, I haven’t even replied in this thread, and you just seem like you’re on the warpath with anyone that wants to defend using LLMs. If Greg KH thinks it’s coming into it’s own, you might want to heed him.
Literally who are you talking about?
Greg Kroah-Hartman (KH) or are you just being obtuse? About whom this article is written and not exactly some rando on the internet like you and me.