Qwen3 4B

Qwen3-4B-FP8 is the FP8-quantized build of the 4 billion parameter model, aimed at on-device and edge inference with a small memory footprint while retaining instruction-following capability. The fine-grained quantization with a 128 block size trims memory needs for smaller hardware. Despite its compact size, the model supports switchable thinking and non-thinking modes: it can generate step-by-step reasoning in think blocks for mathematics, logic, and coding, or answer directly for conversational tasks. It supports a 128K context window for longer documents and code files, along with tool calling for agentic applications. Drawing on the broad pretraining of the Qwen3 family and strong multilingual coverage across 100+ languages, Qwen3-4B competes with larger models from earlier generations on general knowledge and reasoning. It suits mobile devices, embedded systems, rapid prototyping, and latency-sensitive scenarios. Open-weight under Apache 2.0.