Qwen3.6-35B-A3B drops amid secrecy concerns

The 35B-A3B release solidifies Qwen as the leader in "efficient frontier" models, specifically targeting autonomous agent loops.

• –Sparse MoE architecture (35B total, 3B active) enables high-performance inference on consumer-grade dual-GPU setups.
• –Native "thinking preservation" features solve "agent amnesia," allowing stable reasoning across multi-step repository-level tasks.
• –Performance on SWE-bench Verified (73.4%) rivals much larger dense models, proving the scaling laws of sparse routing.
• –Growing community concern suggests a "management" shift in AI labs, potentially prioritizing proprietary APIs over future open-weight releases.