Alibaba announced the launch of Qwen3.8, a large language model featuring 2.4 trillion parameters, on Saturday. The company claims Qwen3.8 performs “second only to Fable 5” overall. A preview version named Qwen3.8-Max-Preview is currently available to developers through Alibaba’s Token Plan subscriptions and its coding platforms Qoder and QoderWork.
This model represents a significant scale increase from prior releases, including Qwen3-Max with 1 trillion parameters launched in September 2025 and Qwen3.5 with 397 billion parameters introduced in February 2026. Qwen3.8 more than doubles the parameter count compared to Qwen3.5 and is described as fully multimodal.
In a departure from its recent practice, Alibaba’s Qwen team stated on social media that open weights for Qwen3.8 would be available “soon.” This strategy contrasts with the previously proprietary status of flagship Max models; for instance, the Qwen3.7-Max model released in May 2026 did not have open weights.
Performance claims for Qwen3.8 originate from internal evaluations, with no independent third-party benchmarks available to verify the assertions. This pattern mirrors prior Qwen releases; for example, Qwen3.5’s performance claims of being comparable to GPT-5.2 and Claude 4.5 Opus were also self-reported and lacked third-party verification.
Competition among AI models between Chinese and American labs increased throughout 2026, with Alibaba releasing several models purporting to compete with those from OpenAI, Anthropic, and Google DeepMind. Currently, developers in mainland China can access Qwen3.8-Max-Preview.
The anticipated release of open weights could revive the strategy that contributed to earlier Qwen models’ widespread adoption in the open-source AI community. Alibaba has not communicated a timeline for the full release of Qwen3.8 nor specified its architecture, training data, or context window details.
