Question for the Llama team: what will be your play? π
Yi Cui
AI & ML interests
Recent Activity
Organizations
onekq's activity


Post train a Qwen model with a dataset distilled from DeepSeek π

Not implying. I like to know what the base is. If QwQ and DeepSeek distill use the same base, then it becomes more puzzling why the performance differ so much.

Ah I see. Thanks!
Still the blogpost didn't mention what the base model is (if any).

Cool! I will check it out.
What I meant by switching is this. Sometimes I'm not satisfied with ChatGPT answer, and realized it needs to think harder. So I switched to o1 and asked again, and most of the times the answer gets better. Then I asked a simple follow-up question which o1 overanalyzed. Then I had to switch back to gpt-4o. I don't actually have the foresight which model fits my question the best. I only know it after I read the answer which is too late.
Now imagine a conversation with a human expert. A human can do such switching remarkably well, hence a cool conversation. This can be actually a metric to read the mileage of an applicant.

This makes it particularly mysterious what went into QwQ-32B? Why did it work so well? Was it trained from scratch? Anyone has insights about this?
onekq-ai/WebApp1K-models-leaderboard

We now have powerful models capable of either system I thinking or system II thinking, but not both, much less switching between the two. But humans can do this quite easily.
ChatGPT and others push the burden to users to switch between models. I guess this is the best we have now.

It ranks below o1-preview, but beats DeepSeek v3 and all Gemini models.
onekq-ai/WebApp1K-models-leaderboard
Now we have such a powerful model that can fit into a single GPU, can someone finetune a web app model to push SOTA of my leaderboard? π€

(1) expensive and even worse, slow, due to excessive token output. You need to 10x your max output length to avoid clipping the thinking process.
(2) you have to filter thinking tokens to retrieve the final output. For mature workflows, this means broad or deep refactoring.
1p vendors (open-source and proprietary) ease these pain points by manipulating their own models. But the problems are exposed when the reasoning model is hosted by 3p MaaS providers.


Then I realized that FP8 should be the main driving force here. So your raw inter-GPU bandwidth is cut in half (H800). But if you compress your data presentation from 16 bits to 8 bits, then the effective throughput of your workload stays unchanged!
The idea is simple but lots of work had to be done. Their v3 technical report will give you a wholistic view (better than reading the code). To summarize, data structure is the foundation to any software. Since FP8 was new and untried, the ecosystem wasn't there. So DeepSeek became the trailblazer. Before cooking your meals, you need to till the land, grow crops, and grind the flour π

This is my summary to πDeepSeekπ open source week. H800 is as good as H100, except the NVLink bandwidth is cut in half.
This is a crystal clear challenge, and it rallied and motivated innovations which follow. The rest are details.

onekq-ai/WebApp1K-models-leaderboard
No SOTA for new models by both OAI and Anthropic. This is not a coincidence. You cannot make everyone happy when more and more workflows and applications use a single model.
Vertical models will inevitably rise.

πDeepSeek π, the first question to ask is why.
The keyword here is H800, a lower-end product tailored for export control. The purpose here is to squeeze out as much performance as possible.
But here is the most important takeaway: this invention benefits EVERYONE.

hi

onekq-ai/WebApp1K-models-leaderboard
I'm sure though this version improves on something, only not the thing my leaderboard measures. This proves the point that no model can be the best on everything.

Done. So I understand this: you do not change model weights, but rather tweak the inference logic? Somehow remind me of speculative decoding.