Post #6638

🚀 Alibaba drops Qwen3.6 — 3B Active, competing with giants Alibaba just open-sourced Qwen3.6-35B-A3B, a sparse MoE model with 35B total params but only 3B active under an Apache 2.0 license. Why it matters: • Agentic coding rivals models 10× larger • Native multimodal (vision + text) reasoning • Supports both thinking + non-thinking modes Performance punch: • Beats its predecessor Qwen3.5-35B-A3B by a wide margin • Outperforms dense 27B models on key coding benchmarks • Matches (and sometimes beats) Anthropic’s Claude Sonnet 4.5 on vision-language tasks • Strong spatial intelligence (e.g. 92.0 on RefCOCO) Open-source + extreme efficiency is serious pressure on closed models. https://huggingface.co/Qwen/Qwen3.6-35B-A3B @aipost🏴