Qwen3.6 MoE narrows dense gap

The interesting story is not that dense still wins; it is that sparse local models are getting close enough that deployment constraints now matter as much as leaderboard deltas.

• –MoE’s 3B active-parameter profile makes the 35B-A3B model much more attractive for consumer-GPU and Mac users chasing speed, context, and memory efficiency.
• –Coding is the pressure point: shrinking gaps on SWE-bench-style evals suggest sparse models are becoming practical for local coding assistants, not just chat demos.
• –Terminal-Bench remains the warning label, because agentic shell work can punish routing instability, tool-use quirks, and long-horizon consistency.
• –The 262K context window is a real differentiator if it works reliably in local stacks, but users should test latency, KV-cache pressure, and quality degradation before assuming it is production-ready.