What's this about?
Alibaba's Qwen team announced Qwen 3.8 — a 2.4-trillion-parameter model that is "launching and going open-weight soon." The tweet claims it is "second only to Fable 5," i.e. better than every other model on Earth except Anthropic's flagship. No weights, benchmarks, or technical report yet — this is a teaser, dropped three days after Moonshot AI announced its own 2.8T open model Kimi K3.
What exactly does "open-weight" mean?
A model's weights are the trained numbers that make up its "brain" — the file you'd need to actually run it yourself. "Open-weight" means that file gets published for anyone to download, like releasing the master recording instead of only selling streaming access. It doesn't necessarily include the training data or code, so it's "open" in a narrower sense than open-source software. At 2.4T parameters, though, running it yourself means a serious GPU cluster, not a laptop.
What's the community saying?
- This is a Kimi K3 response — the top comment reads the timing as Alibaba reacting to Moonshot's July 27 weights promise, with GLM 5.2 and Xi Jinping's recent open-source commitment at the Shanghai AI conference as extra fuel: "from this competition in LLMs, we win."
- "Second only to Fable 5" is quite an admission — HN notes that even the challengers now benchmark themselves against Anthropic's model as the ceiling; whatever moat debate existed, the challengers themselves have settled it for now.
- Skepticism about Qwen specifically — testers call Qwen the most censored of the Chinese model families and report Qwen 3.7 Pro being far behind DeepSeek V4 in real coding work; "I predict everyone will use Kimi K3 instead."
- The typo discourse — "compatible to leading frontier AI models" instead of "comparable": how does a frontier lab not proofread its own launch tweet?
- Please ship small ones too — local-inference users are begging for 35B/80B versions they can actually run at home.