Qwen3-0.6B

Qwen3-0.6B is the smallest dense model in Alibaba's Qwen3 family, released April 29, 2025 under the Apache 2.0 license. Despite its tiny size, it inherits the family's hybrid architecture — switchable between 'thinking' mode for multi-step reasoning and fast non-thinking mode — plus support for 119 languages and dialects. It was trained as part of the same 36-trillion-token run as the rest of the Qwen3 lineup, which spans from this 0.6B model up to the 235B-A22B MoE flagship. A 4-bit MLX build exists for native Apple Silicon execution, making it one of the easiest ways to run a reasoning-capable model on a base-model Mac.\n\nThe honest picture: at 0.6B parameters, world knowledge is thin and reasoning depth is bounded — this is a model for lightweight agents, on-device assistants, and embedded tasks, not a general-purpose problem solver. But as a fully permissive, sub-billion reasoning model, it fills a niche that few other labs bother to serve.