Qwen3-4B-Thinking-2507

Qwen3-4B-Thinking-2507 is the smallest reasoning model in the Qwen3-2507 refresh, bringing dedicated chain-of-thought capability to a 4-billion-parameter dense model that can run on a single consumer GPU. Announced alongside the 235B-A22B and 30B-A3B thinking variants in July 2025, it continues the Qwen3 thinking line with improved reasoning quality and depth over the original hybrid release, while keeping the 256K native context window extendable to 1 million tokens.

The 4B thinking model is notable because it makes serious reasoning workloads viable on edge hardware and in cost-sensitive deployments. It ships under Apache 2.0 on Hugging Face as Qwen/Qwen3-4B-Thinking-2507 and works with the standard Qwen3 inference stack including Transformers, vLLM, and llama.cpp quantizations. Per-size dates within the July 2025 refresh are era-level.