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)

Qwen3-Coder and its Qwen2.5-Coder predecessor both ship Apache-2.0 weights with no use restrictions, but the training corpus is closed and only inference code and a chat template are released, with no training pipeline. That keeps the family at open weights rather than open source, across both generations.

• https://huggingface.co/api/models/Qwen/Qwen3-Coder-480B-A35B-Instruct recorded 2026-09-29

  Hub API record for "Qwen/Qwen3-Coder-480B-A35B-Instruct" carries the tag "license:apache-2.0", with safetensors weights and the repo not gated.

• https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct recorded 2026-09-29

  Model card lists "Training Stage: Pretraining & Post-training" and covers quickstart and agentic-coding usage; searched the body for training corpus, post-training data, SFT recipe and training pipeline: none released.

Adoption is measured as Hugging Face downloads summed across the four declared Qwen Coder SKUs, spanning both the 2.5 and 3 generations. Most of that volume comes from the older Qwen2.5-Coder SKUs, and the largest single SKU alone draws enough downloads to support the same reading as the four combined.

• https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct recorded 2026-08-13

  1,242,001 downloads in the trailing 30 days

Qwen's model card says the model achieves results comparable to Claude Sonnet on agentic coding among open models; the claim is vendor-reported and no independent SWE-bench result is cited. It sits a step below current frontier coding models.

• https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct recorded 2026-08-13

  vendor claim of results "comparable to Claude Sonnet" on agentic coding and agentic browser-use among open models; 480B-A35B MoE with 256K native context extendable to 1M via Yarn