Qwen3.6-35B-A3B

The sparse MoE architecture here is the real move — 35B total params but only 3B active means it punches way above its weight class in coding tasks while actually being computationally efficient, unlike those dense giants that eat GPU for breakfast. It's Apache 2.0 licensed so it's genuinely free (not some "open" model hiding behind API walls), and the dual thinking/non-thinking mode is a smart touch for cases where you don't need full chain-of-thought overhead. The traction is solid for a day-old launch — #13 with 2.1K followers and JDoodle.ai already building on it — and seeing real developers actually use it matters more than vanity metrics.