Qwen3.6 bets on coding agents and persistent reasoning

What it does

Qwen3.6 is the latest open-weight release from Alibaba’s Qwen team, building on Qwen3.5 with a focus on stability and real-world developer productivity. It emphasizes agentic coding—handling front-end workflows and repository-level reasoning—and introduces “thinking preservation,” which retains reasoning context across conversation history. The initial release includes a 27B dense model and a 35B-A3B sparse MoE, both under Apache 2.0.

The interesting bit

The thinking preservation feature is a pragmatic UX pivot: instead of discarding its work-through between turns, the model keeps it, aiming to reduce overhead during iterative coding sessions. Under the hood, Qwen3.5 provides the foundation with sparse Mixture-of-Experts architectures, vision-language fusion trained on trillions of tokens, and reinforcement learning scaled across million-agent environments.

Key highlights

• Open weights under Apache 2.0, hosted on Hugging Face and ModelScope with inference support via vLLM, SGLang, llama.cpp, and Apple Silicon MLX
• Qwen3.6 targets agentic coding and preserves thinking context across conversation history
• The Qwen3.5 foundation adds a 397B-A17B MoE, Gated Delta Networks, multimodal training at near-100% efficiency, and coverage for 201 languages
• Ecosystem includes Qwen Studio (chat UI), Qwen Code (terminal agent), and the Qwen-Agent framework
• Alibaba Cloud Model Studio provides OpenAI- and Anthropic-compatible API access

Caveats

• User guide and documentation are marked “coming soon” in multiple places
• This repository serves primarily as an information hub; model weights and implementation code reside on external platforms

Verdict

Worth evaluating if you need an Apache 2.0 coding model with persistent reasoning memory. Those requiring complete documentation today or a self-contained source-code release should look elsewhere.