> The training and deployment of LongCat-2.0 are built on large-scale clusters of tens of thousands of AI ASIC superpods. Compared to the mature Nvidia GPU ecosystem, the supporting software community is still less developed. We have therefore put significant effort into building a stable, secure, and scalable infrastructure.
Dwarkesh Patel has AI/ML guests on his podcast. BoorishBears may have been referring to the Jensen Huang episode where they discuss TPUs: https://youtu.be/Hrbq66XqtCo?t=982
> If you could run a nuclear reactor with U-235 as fuel or Pu-241 (both mixed with 95% U-238), which one would you choose and why?
For a human this would not be tricky at all. For an LLM it could be, because this question certainly does not exist in any sort of training, because Pu-241 does not exist in pure form, it only exist as a minor component of reactor-grade plutonium, where Pu-239 would dominate, with Pu-240 coming second and Pu-241 coming third.
In any case, LongCat-2.0. gave a very well reason but incorrect answer that Pu-241 is preferable.
I then tested on Qwen 3.7 Plus, and it correctly answered that U-235 is preferable because of its much higher delayed neutron fraction. I then went to Gemini Flash, which answered the same, with much more confidence, and with much stronger arguments, and the speed of the answer was much higher.
Overall I rate Gemini Flash the best, Qwen 3.7 Plus an acceptable second, and LongCat-2.0 an ok'ish third, if you have nothing better.
I am not a physicist but perhaps your question was leading more than you expected? I would take the question to pre-suppose I have an abundance of the stated material, ignoring practical realities of refinement. If I did have fully pure Pu-241, would that be a better fuel than U-235?
Or stated another way, "If you could run a generator on gasoline or jet fuel, which one would you choose and why?" I would answer jet fuel owing to slightly higher energy density and purity of the material - likely leading to a cleaner burn. Which would ignore that jet fuel is going to be a multiple of the gasoline price.
If I did have fully pure Pu-241, would that be a better fuel than U-235?
Also not a physicist, but I assume from the fact that the OP is asking the LLM this question to trip it up, the point is that U-235 is better even if you have an abundance of both. It's scarcity of Pu-241 leads to the lack of data in training, not that it's actually better.
Again, really speaking out of my depth, but if there is a lack of plutonium training data, I would assume the answer would be the far more commonly described U-235. To respond otherwise means there is some existing association with Pu-241 being better.
I don't think this is where you were going with your comment, but I'll mention this just because you're somewhat adjacent to a routine mistake in business:
Uber is a people delivery company, but they've had a lot of bright engineers working for them on their infrastructure and software over the years, and that work has rippled out through the industry.
Amazon (in VMWare's words) is "a company that sells books", and their leadership couldn't accept they were losing to them ("I look at this audience, and I look at VMware and the brand reputation we have in the enterprise, and I find it really hard to believe that we cannot collectively beat a company that sells books.").
The thing that stood out for me about Meituan was that their power bank rental gizmos were everywhere in China, and people would rather rent a power bank than own and carry one around because of how convenient it is.
I asked a question with "Search" enabled, with the app set to English, and got results back in Chinese. Interesting view into how the LLM responds to its context.
I don't think llama.cpp supports any of the LongCat models, actually.
They haven't posted weights/inference solutions for LongCat-2.0 [1], but LongCat-Next had transformers support, which I assume means it works with vLLM/SGLang.
Given it's 1.6T, "common hardware" is probably out of the question; even 2bpw is going to measure out at 400GB, even before considering the bandwidth requirements for 48B active. I haven't read the LongCat-2.0 architecture docs, but if you're not running GLM-5.2, you're probably not running this either.
Yeah, for me it seems like a if you have to ask you can't run it" type question.
In general the TL;DR is that anything above 35B needs hardware you buy basically only to run large LLMs, and if you have that hardware you don't need to ask the question.
~70B models can run fine (albeit somewhat slow) on consumer hardware with 64GB RAM. There are heavily quantized (Q1.x) models that are still usable on similar hardware. Granted recently there haven't been a lot of models of this size, but still, 35B isn't really the practical limit. 35B is mostly the limit if you're using consumer grade GPUs with limited RAM and need the model to run fast.
People have been toying with running large-ish models by partially offloading on CPU+RAM with mixed results, but as long as you're OK with reduced speed, and you quantize the hell out of the big models, you can apparently try a lot more models locally than popular belief.
Ah yes but because it’s a MoE 48GB active model, then it’s possible that we might be able to run it locally in specialised setups such as 256GB unified memory.
Many MoE models (seem?) to only require enough memory to load the active expert.
So... is this literally a... umm, sorry, I'm just genuinely (really, no sarcasm intended) which terminology to use... finetune of DeepSeek V4-Pro or post-trained version of DeepSeek V4-Pro Base? Because I haven't fully dived into the tech report (so I may update my opinion as well as my comment), but this far the architectural solutions seem to be largely similar to DeepSeek ones.
Maybe I'm wrong, but that's just the first impression.
EDIT: I take my words back (which happens rarely) - although they do build upon DeepSeek's work, their contribution far exceeds merely post-training the base model in a different way. They did introduce something new to the architecture, though I still can't find the full tech report, with Hugging Face and GitHub links returning 404 right now.
EDIT-2: Now when I think about it, I'm not quite sure if they're going to release in the open the full report with methodology, as well as the model weights, at all.
If more people are doing what DeepSeek did and figuring it out, that's a great thing, because DeepSeek figured out how to radically reduce the cost of inference.
This is the real news story. It looks like they may have used Huawei Ascend 910C chips: https://nitter.net/teortaxesTex/status/2071708141037781407#m
In any case, LongCat-2.0. gave a very well reason but incorrect answer that Pu-241 is preferable.
I then tested on Qwen 3.7 Plus, and it correctly answered that U-235 is preferable because of its much higher delayed neutron fraction. I then went to Gemini Flash, which answered the same, with much more confidence, and with much stronger arguments, and the speed of the answer was much higher.
Overall I rate Gemini Flash the best, Qwen 3.7 Plus an acceptable second, and LongCat-2.0 an ok'ish third, if you have nothing better.
Or stated another way, "If you could run a generator on gasoline or jet fuel, which one would you choose and why?" I would answer jet fuel owing to slightly higher energy density and purity of the material - likely leading to a cleaner burn. Which would ignore that jet fuel is going to be a multiple of the gasoline price.
Also not a physicist, but I assume from the fact that the OP is asking the LLM this question to trip it up, the point is that U-235 is better even if you have an abundance of both. It's scarcity of Pu-241 leads to the lack of data in training, not that it's actually better.
/s
Uber is a people delivery company, but they've had a lot of bright engineers working for them on their infrastructure and software over the years, and that work has rippled out through the industry.
Amazon (in VMWare's words) is "a company that sells books", and their leadership couldn't accept they were losing to them ("I look at this audience, and I look at VMware and the brand reputation we have in the enterprise, and I find it really hard to believe that we cannot collectively beat a company that sells books.").
A bonus would be tok/s on common hardware.
They haven't posted weights/inference solutions for LongCat-2.0 [1], but LongCat-Next had transformers support, which I assume means it works with vLLM/SGLang.
Given it's 1.6T, "common hardware" is probably out of the question; even 2bpw is going to measure out at 400GB, even before considering the bandwidth requirements for 48B active. I haven't read the LongCat-2.0 architecture docs, but if you're not running GLM-5.2, you're probably not running this either.
[1] https://huggingface.co/meituan-longcat/LongCat-2.0: "Model weights coming soon — stay tuned!"
In general the TL;DR is that anything above 35B needs hardware you buy basically only to run large LLMs, and if you have that hardware you don't need to ask the question.
~70B models can run fine (albeit somewhat slow) on consumer hardware with 64GB RAM. There are heavily quantized (Q1.x) models that are still usable on similar hardware. Granted recently there haven't been a lot of models of this size, but still, 35B isn't really the practical limit. 35B is mostly the limit if you're using consumer grade GPUs with limited RAM and need the model to run fast.
People have been toying with running large-ish models by partially offloading on CPU+RAM with mixed results, but as long as you're OK with reduced speed, and you quantize the hell out of the big models, you can apparently try a lot more models locally than popular belief.
Many MoE models (seem?) to only require enough memory to load the active expert.
Maybe I'm wrong, but that's just the first impression.
EDIT: I take my words back (which happens rarely) - although they do build upon DeepSeek's work, their contribution far exceeds merely post-training the base model in a different way. They did introduce something new to the architecture, though I still can't find the full tech report, with Hugging Face and GitHub links returning 404 right now.
EDIT-2: Now when I think about it, I'm not quite sure if they're going to release in the open the full report with methodology, as well as the model weights, at all.