Intern-S1-Pro (1T/A22B)

Posted by ResearchCrafty1804@reddit | LocalLLaMA | View on Reddit | 26 comments

🚀Introducing Intern-S1-Pro, an advanced 1T MoE open-source multimodal scientific reasoning model. \- SOTA scientific reasoning, competitive with leading closed-source models across AI4Science tasks. \- Top-tier performance on advanced reasoning benchmarks, strong general multimodal performance on various benchmarks. \- 1T-A22B MoE training efficiency with STE routing (dense gradient for router training) and grouped routing for stable convergence and balanced expert parallelism. \- Fourier Position Encoding (FoPE) + upgraded time-series modeling for better physical signal representation; supports long, heterogeneous time-series (10\^0–10\^6 points). \- Intern-S1-Pro is now supported by vLLM @vllm\_project and SGLang @sgl\_project @lmsysorg — more ecosystem integrations are on the way. Huggingface: https://huggingface.co/internlm/Intern-S1-Pro GitHub: https://github.com/InternLM/Intern-S1

26 Comments

[-]

pulse77@reddit

Can someone discover a good MoE architecture which will select those A22B and leave the rest on SSD (most of the time) so this could run fully in 24GB VRAM - even without RAM?

Karyo_Ten@reddit

>Can someone discover a good MoE architecture which will select those A22B and leave the rest on SSD (most of the time) so this could run fully in 24GB VRAM - even without RAM? You can just create a 1TB swapfile.

crantob@reddit

Why downvote? This is actually doing what the requester asked.

Former-Ad-5757@reddit

So basically any dense 22b model. The power of moe’s above dense models is the fact that every token can get rerouted different in the 1 tb model, you can do that currently with llama.cpp, it will just run at 1 token a day because every token needs to read from disk, there is no 22b which can stay in memory and is reused,

With 24GB VRAM + 128GB RAM + SSD it is about 1 token/s (tested on my machine).

Lissanro@reddit

Good MoE architecture does the exact opposite, ideally during long inference on average all experts should get used equally. In practice some may be "hotter" than others. There were also attempts like Reap to cut down least important experts but this always leads to lesser quality and reduction of knowledge, especially in less popular areas.

sine120@reddit

I like these specialist models. I don't personally have any use for it, but it's cool to see the segmentation, since not all models need to do everything.

Also please no AI designed bioweapons, thanks

bene_42069@reddit

Like anything, it's a double edged sword. Whether it will do good or harm depends on the trainer/user/owner.

Someone right now is using AI to kill people but people seem too shy to name the evil.

InternationalNebula7@reddit

Wow. Now I just need my own personal data center to run it.

Healthy-Nebula-3603@reddit

Or just HEDIT machine with 12 channels of DDR5 and one TB .

I have to wait for llama.cpp support before I can try it. In the mean time, I will keep using K2.5 (Q4_X quant). But Intern-S1-Pro looks very interesting because has 22B active parameters instead of 32B like K2.5, so potentially can be faster.

VoidAlchemy@reddit

Yeah AesSedai's K2.5 Q4\_X is probably best available open quant now, though 22B active parameters here with Intern-S1-Pro sounds promising for speed assuming it is any good and gets a solid implementation.

Aggressive-Bother470@reddit

I might start a gofundme for 1TB RAM.

pigeon57434@reddit

buy 3,090 3090s and youll be set to run this baby in full precision

JustSayin_thatuknow@reddit

You can add a tool call to allow your model to use the calculator, please don’t rely on LLMs for math stuff 😆

That math of yours.. oh man😜

lan-devo@reddit

Help Aggressive-Bother470 to overcome his addiction by fulfilling it to the max

ZestyCheeses@reddit

I would happily pay into some sort of fund that purchases physical compute. Then we all vote on what we want the AI to do or research etc. I don't see how the common man will be able to keep up with large capital holders unless we build compute unions of some sort.

Intern-S1-Pro (1T/A22B)

Reply to Post

26 Comments

pulse77@reddit

Karyo_Ten@reddit

crantob@reddit

Former-Ad-5757@reddit

pulse77@reddit

Lissanro@reddit

sine120@reddit

sine120@reddit

bene_42069@reddit

crantob@reddit

InternationalNebula7@reddit

Healthy-Nebula-3603@reddit

Lissanro@reddit

VoidAlchemy@reddit

Aggressive-Bother470@reddit

pigeon57434@reddit

JustSayin_thatuknow@reddit

JustSayin_thatuknow@reddit

lan-devo@reddit

ZestyCheeses@reddit

Daemontatox@reddit

Karyo_Ten@reddit

_Anime_Anuradha@reddit

Alternative-Theme885@reddit

Signature97@reddit

SlowFail2433@reddit