Alibaba publishes Qwen3.8 open models focused on long-horizon agentic coding
AI agents · Friday, 2 October 2026
Why it matters
Open availability plus support across major inference and serving stacks lowers the cost of testing Qwen3.8 as an agent backend or coding-agent foundation. The explicit controls for reasoning depth and retained context are relevant to product designs that need to balance long-horizon task performance against latency and inference cost.
What happened
Alibaba’s Qwen team published the Qwen3.8 repository and made Qwen3.8-2.4T-A95B and Qwen3.8-27B available through Hugging Face Hub and ModelScope. The release targets coding, professional work, research, and complex multi-step agentic workflows, with features for autonomous planning, environment feedback, adjustable reasoning depth through `reasoning_effort`, and retained reasoning context through `preserve_thinking`. Developers can run or deploy the models through tools including Transformers, llama.cpp, MLX, Unsloth, SGLang, vLLM, and TokenSpeed, while Alibaba also lists access through Qwen Studio, Qoder, QwenWork, QwenCloud, and Qwen Code.
Players & places
- Alibaba
- Qwen
- Hugging Face
- ModelScope