Alibaba releases weights for Qwen3.8-2.4T-A95B large language model

Alibaba has announced the open release of its Qwen3.8-2.4T-A95B large language model weights through its ModelScope community. This marks the first time weights for a Qwen-Max level model have been made publicly available.

The model boasts a total of 2.4 trillion parameters, activating 95 billion parameters per token. Notably, it natively supports a context window of 262,144 tokens, which can be extended to over one million tokens. This significant context capacity allows for the processing and understanding of much longer texts and dialogues.

Qwen3.8 is specifically engineered to enhance performance in programming, office tasks, scientific research, and long-duration Agent tasks. Alibaba's cloud version, Qwen3.8-Max, is built upon these released weights and incorporates additional capabilities for production environments.

Initial evaluations suggest the model is competitive with other leading large language models, including GPT-4 and Opus. It has demonstrated strong performance across various benchmarks, including those for programming Agent tasks, general Agent capabilities, professional work, and long-context understanding. The model's architecture utilizes Mixture-of-Experts (MoE), with 512 experts per layer, and 10 experts activated per token.