Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All

What Qwen released

Qwen has open-sourced Qwen3.6-35B-A3B, a mixture-of-experts (MoE) model that packs 35 billion total parameters but activates only about 3 billion at inference. It has been reported that despite the sparse footprint, the model delivers “exceptional agentic coding capability” and supports both multimodal and non-thinking modes. In plain English: smaller active compute, big-league behavior. Why should you care? Because open weights mean researchers and hobbyists can poke, prod, and build without waiting for gated APIs.

Benchmarks and performance

Qwen’s blog presents head-to-head numbers showing the new MoE beating its direct predecessor, Qwen3.5-35B-A3B, and matching or outpacing much larger dense models like Qwen3.5-27B and Gemma-31B on several coding and agent benchmarks. It has been reported that Qwen3.6-35B-A3B scored 51.5 on Terminal-Bench 2.0 (vs. 40.5 for Qwen3.5-35B-A3B) and posted a QwenWebBench rating of 1397, comfortably ahead of peer entries. Those figures come from Qwen’s internal and public evaluations (SWE-bench, Claw-Eval, SkillsBench, NL2Repo and others), with experimental details provided alongside the results.

Availability

The model is available for interactive use on Qwen Studio, accessible through Qwen’s API (listed as Qwen3.6-Flash on some clouds; Alibaba Cloud Model Studio API support is coming soon), and distributed as open weights on Hugging Face and ModelScope. That’s the full stack: try it in the browser, call it from your app, or download the weights and run it locally if you’ve got the chops.

Why it matters

Agentic coding is the hot ticket right now — autonomous agents that can edit files, run terminals, and stitch together multi-step workflows are reshaping developer tooling. Open-sourcing a high-performing MoE at this efficiency point nudges the whole field toward cheaper, more capable models that teams can experiment with without clearing corporate hoops. Call it democratization with a turbocharger: more people get the tools, and we’ll find out fast where it shines and where it stumbles.

Sources: qwen.ai, Hacker News