Alibaba’s Qwen3.8 open models focus on agentic coding

Alibaba’s Qwen team has published the GitHub repository for Qwen3.8, the latest entry in the Qwen3.5 open model series alongside Qwen3.5 and Qwen3.6. The release brings a Qwen-Max-class model to open availability and emphasizes gains in coding, professional work, research and long-horizon agentic tasks, with a focus on completing complex multi-step workflows more reliably.

Qwen3.8 adds stronger autonomous planning, improved handling of environment feedback and broader support for common harnesses and development tools. Developers can tune reasoning depth through reasoning_effort and preserve reasoning context from prior messages with preserve_thinking, while official access is available through Qwen Studio, Qoder, QwenWork, QwenCloud, Qwen Code, Hugging Face Hub and ModelScope.

The repository lists Qwen3.8-2.4T-A95B and Qwen3.8-27B as available on Hugging Face Hub and ModelScope, with local and deployment paths through transformers, llama.cpp, MLX, Unsloth, SGLang, vLLM and TokenSpeed. Earlier Qwen3.5 updates highlight multimodal training, hybrid architecture, reinforcement learning scale and support for 201 languages and dialects.