Alibaba announces 2.4T parameter Qwen3.8-Max model

Alibaba has officially unveiled Qwen3.8-Max, a massive flagship model boasting 2.4 trillion parameters built on a sparse mixture-of-experts architecture. The tech giant claims the model ranks second globally, trailing only Claude Fable 5. Alongside this release, Alibaba promised to open-source the weights for both Qwen3.8-Max and a smaller Qwen3.8-27B model next week. Currently, the flagship model is accessible as a preview through Qwen Studio and its API, priced at $2.0 per million input tokens and $6.0 per million output tokens. It features a massive 983,616-token context window and a 131,072-token maximum output limit.

Before its official debut, the model made a stealthy appearance on the Code Arena leaderboard under the pseudonym kaleb. It initially identified itself as Claude due to a distillation training artifact, but community members quickly unmasked it by identifying a unique PostalCodesNL token pattern from the Qwen tokenizer. On the Code Arena leaderboard, Qwen3.8-Max leads the 2.8-trillion-parameter Kimi K3 model by approximately 6 Elo points. In autonomous coding demonstrations, the model successfully constructed the oh-my-cli repository over 16 days, generating 422 commits with zero tool-call failures during independent testing.

For AI practitioners, this release represents a major shift in the open-weights landscape, which has recently seen top-tier Max models remain closed. Having access to a 2.4-trillion-parameter model with such a massive context window could democratize high-end agentic workflows and complex coding tasks. However, developers should remain cautious: there are currently no independent benchmark scores from organizations like Artificial Analysis, Arena.AI, or Hugging Face, meaning all performance claims currently rely on Alibaba's internal evaluations.