Qwen3.6-35B-A3B: Agentic coding power, now open to all
Qwen released Qwen3.6-35B-A3B as open weights, with 35B total parameters and 3B active parameters. The post reports 73.4 on SWE-bench Verified, 51.5 on Terminal-Bench 2.0, and 92.0 on RefCOCO. The key point is agentic coding and multimodal performance at a 3B active-parameter budget, with weights, Qwen Studio, and API access available.
Why it matters: Qwen open-sources Qwen3.6-35B-A3B, a coding-agent model with 35B total parameters that activates only 3B per inference, so the trade-off is deployment cost.