Qwen3.6 27B, 35B-A3B upend 30B models

The hot take: for coding-first local workflows, Qwen3.6 is probably the new reference point, but "obsolete" is too strong unless your only metric is benchmark score.

• –Qwen’s own release notes say Qwen3.6-27B beats the previous open-source flagship Qwen3.5-397B-A17B on major coding benchmarks, which is a brutal result for a dense 27B model.
• –Qwen3.6-35B-A3B is a sparse MoE checkpoint with only 3B active parameters, so it changes the deployment math: better capability than older ~30B models, with much better efficiency than a fully active dense model of similar size.
• –The older ~30B models still have niches: some are faster at a given quantization, some are better tuned for specific instruction styles, and some teams will prefer them for stability, latency, or existing fine-tunes.
• –For agent workflows, raw benchmark leadership is only half the story; tool calling reliability, context preservation, and frontend/repo reasoning matter more than a leaderboard win.
• –If you are choosing one default local model today, Qwen3.6-27B is an easy candidate. If you need throughput per watt or per GPU, Qwen3.6-35B-A3B is the more interesting systems play.