Alibaba’s Qwen team just released QwQ-32B-Preview, a powerful new open-source AI reasoning model that can reason step-by-step through challenging problems and directly competes with OpenAI’s o1 series across benchmarks.
The details:
QwQ features a 32K context window, outperforming o1-mini and competing with o1-preview on key math and reasoning benchmarks.
The model was tested across several of the most challenging math and programming benchmarks, showing major advances in deep reasoning.
QwQ demonstrates ‘deep introspection,’ talking through problems step-by-step and questioning and examining its own answers to reason to a solution.
The Qwen team noted several issues in the Preview model, including getting stuck in reasoning loops, struggling with common sense, and language mixing.
Why it matters: Between QwQ and DeepSeek, open-source reasoning models are here — and Chinese firms are absolutely cooking with new models that nearly match the current top closed leaders. Has OpenAI’s moat dried up, or does the AI leader have something special up its sleeve before the end of the year?
It’s only open source if the training data is and it probably isn’t, is it?
I don’t know, though DeepSeek talk of theirs being “fully” open-source.
Part of the advantage of doing this (apart from helping bleed your rivals dry) is to get the benefit of others working on your model. So it makes sense to maximise openness and access.
Realistically, no LLM that’s large enough to be competitive will be able to remain open-source, even if it was initially (and most that claim to be weren’t actually, as you point out), because so much training data is needed.
Often the training data can’t be re-distributed in the first place, but even if it can be, its availability makes it much more likely that someone will request the takedown of some data in the set (even if the data was licensed, someone who holds copyright might claim that the person who submitted it to the set wasn’t permitted to do so). At that point, unless the takedown request is refused or the model itself is re-trained (which would be quite expensive) the data is no longer sufficient to generate the model.