Qwen3-Coder

weights

open(Apache-2.0, on HF)

data

closed(no training corpus released)

code

partial(inference/usage + chat template

license

Apache-2.0(OSI permissive, no use restrictions)

Apache-2.0 weights with a closed corpus and inference code only, so 3, consistent across both generations - Qwen2.5-Coder and Qwen3-Coder are both plain Apache with no use restrictions. Qwen3-Coder governs and contributes capability 4; adoption 4 is carried from Qwen2.5-Coder, which has had longer to accumulate downloads. Kept apart from the base Qwen tier because Alibaba ships and markets the Coder line separately.

• https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct recorded 2026-06-04

  Apache-2.0 license, 480B-A35B MoE instruct model card, downloadable BF16 weights

~1.36M downloads last month for the 32B-Instruct SKU alone on HF; the broader Qwen2.5-Coder family (0.5B-32B base+instruct) multiplies this. Squarely in the 1-10M reach band on measured monthly download volume.

• https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct recorded 2026-06-04

  ~1,360,411 downloads last month

Strong open-weight coding model (Claude Sonnet-class agentic coding per vendor card), a notch below 2026 frontier coders (DeepSeek-V4-Pro/Kimi K2.6/GLM-5.1). HF card is vendor-reported; no independent SWE-bench number captured this run, so capped at 4 with medium confidence.

• https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct recorded 2026-06-04

  vendor claim of Claude-Sonnet-class agentic coding; MoE architecture and long-context coding focus