Qwen3-0.6B

Qwen3-0.6B is a 600-million-parameter dense causal language model from Alibaba's Qwen team, released on April 29, 2025 under the Apache 2.0 license. It is the smallest model in the Qwen3 family, built on a 28-layer Transformer with Grouped-Query Attention (16 query / 8 KV heads), RoPE, SwiGLU activations, and RMSNorm, supporting a 32,768-token context window. A defining feature is hybrid thinking mode: the model can switch between a default reasoning mode (with chain-of-thought in thinking tags) and a fast non-thinking mode for efficient dialogue, controlled via enable_thinking or /think and /no_think tokens.

On standard benchmarks from the Qwen3 Technical Report, the base model scores 52.81 on MMLU (5-shot), 59.59 on GSM8K (4-shot CoT), and 36.23 on EvalPlus (0-shot coding composite). After post-training and Strong-to-Weak Distillation, the full Qwen3-0.6B model reaches 55.6 MMLU-Redux and 77.6 MATH-500 in thinking mode, outperforming larger baselines such as DeepSeek-R1-Distill-Qwen-1.5B on several reasoning tasks despite having fewer parameters. It also supports 100+ languages, tool calling via Qwen-Agent, and deployment through transformers, vLLM, SGLang, Ollama, and llama.cpp.

Community reception positions Qwen3-0.6B as a leading sub-1B open model for edge and on-device use. Absolute scores lag far larger models (Artificial Analysis Intelligence Index ≈1.0), but within its size class it shows strong gains over Qwen2.5-0.5B and Gemma-3-1B. It is widely used for local inference, agent prototyping, and resource-constrained deployments where latency and memory matter more than frontier-level capability.