Alibaba teases Qwen3.8, a 2.4 trillion parameter model, ahead of open weight release

That's a massive model. The shift from cheap value models to huge, slow models coming out of China is an interesting change in strategy. GLM 5.2 and Kimi 3 already feel token hungry and slow to use, so I hope Qwen doesn't go the same way.

This shift isn't new, Kimi K2 was already a 1T model back in July last year. Glad more labs are following that trend since competitive open models matter.

Value models aren't going away, you can always distill from a bigger model. Having one really smart flagship builds brand confidence, that's part of why the US labs are still holding on.

I bet this announcement was timed to Moonshot AI's Kimi K3, a 2.8T open weight model dropping on HuggingFace July 27. Whatever the motive, this competition between Alibaba and Moonshot is a win for us.

Hard to know their real motivation, but Chinese firms seem intent on commoditizing intelligence, which happens to debase the American frontier labs and is also just good for everyone else.

Don't overthink the timing, when a research area is this hot you get simultaneous releases all the time, that's just how frontier research works.

GLM 5.2's release is probably part of why Alibaba pushed this out too.

I just want a 35B or 80B MoE model I can actually run locally, throw us a bone, not everything needs to be SOTA-sized.

There was a big AI conference in Shanghai recently where Xi Jinping talked up open source AI commitments, so Alibaba rushing this out isn't a surprise.

I'd rather see 3.7-27B or 3.7-122B releases, Qwen and QwQ were always about giving the best local inference you could run at home.

Qwen is way better than GLM or Kimi in my experience, this news genuinely excites me.

Bigger models are usually worse in practice though, hope they actually nail it this time.

The 'second only to Fable 5' line is telling. Anthropic really does have a moat with that model right now, and it'll be a big deal when an open model finally beats it.

Even if it's technically better, Fable's restrictions make it less useful day to day than Sol for my work.

I'd rate the top models Fable > K3 > Sol for my own use, but Fable isn't so far ahead that I'd be hurt without it. It's also the only one of the three that regularly triggers refusals.

I ran a few quick tests and Kimi felt like the real deal, but Qwen seemed more like a benchmark princess to me.

Qwen3.6 is still the best agentic open weight model around 30B params in my experience, and it's noticeably less glitchy if you make it think in Chinese via the system prompt.

I use the 35B MoE and 27B dense Qwen models locally and barely need Claude most days, especially for sensitive or personal data.

A 2.4T model is a completely different architecture, you can't just shrink that down to 35B and keep the accuracy.

Qwen feels like the most censored of the Chinese models when I test it, and DeepSeek V4 Pro beats Qwen 3.7 Max on speed, cost, and quality in my experience.