Alibaba Cloud Open-Sources Qwen3.8-2.4T Model Weights

On August 13, Alibaba Cloud’s ModelScope community announced that the Qwen team has released the model weights for Qwen3.8-2.4T-A95B, marking the first open-source release of weights for a Qwen-Max-level model. This model has a total of 2.4 trillion parameters, with 9.5 billion parameters activated per token, supports a context length of up to 262,144 tokens, and can scale to 1,010,000 tokens. Building upon Qwen3.5’s hybrid architecture, Qwen3.8 focuses on enhancing end-to-end performance in programming, office productivity, scientific research, and long-horizon agent tasks. Leveraging the open weights, Qwen3.8-Max integrates additional capabilities tailored for production environments. Official evaluations show that Qwen3.8 demonstrates competitive performance against models such as Opus4.8, Fable5, and GPT5.6Sol across programming agents, general-purpose agents, professional workflows, and long-context understanding, achieving strong results on benchmarks like PaperBench and IFBench. The post-training process for Qwen3.8-Max in office and agent scenarios comprises three key components: continuous expansion of real-world environments, construction of a unified reward system, and development of an online data balancer—collectively enabling stable and scalable reinforcement learning training.