Shadowfetch News — AI news. Real coverage.

Models

Qwen puts Qwen3-Coder on the agentic coding track

The Qwen3-Coder family pairs an Apache-2.0 model card, long-context coding claims and official deployment metadata for agent-style developer workflows.

a computer monitor sitting on top of a desk
Photo by Boitumelo on Unsplash

Qwen says it is announcing Qwen3-Coder as its most agentic code model to date, and the Qwen3-Coder-480B-A35B-Instruct model card says that variant is the first and most powerful one introduced for the family.

The Hugging Face model card lists Qwen3-Coder-480B-A35B-Instruct under the Qwen organization with an Apache-2.0 license and a text-generation pipeline tag.

The model card says Qwen3-Coder-480B-A35B-Instruct has 480B total parameters, 35B activated parameters, 62 layers, 96 query attention heads, 8 key-value attention heads, 160 experts and 8 activated experts.

The same model card lists a native context length of 262,144 tokens, and Qwen's release text says the long-context path supports 256K tokens natively and can extend to 1M tokens with Yarn.

Qwen's GitHub repository says the Qwen3-Coder family includes Qwen3-Coder-480B-A35B-Instruct, Qwen3-Coder-30B-A3B-Instruct and Qwen3-Coder-Next, and it describes Qwen3-Coder-Next as an open-weight model for coding agents and local development.

The GitHub repository says Qwen3-Coder supports 358 coding languages and includes examples for fill-in-the-middle coding, website release work, desktop cleanup and game-building tasks.

ModelScope's Qwen page identifies Qwen3-Coder-480B-A35B-Instruct as Qwen/Qwen3-Coder-480B-A35B-Instruct, lists Qwen3MoeForCausalLM as the architecture, lists apache-2.0 as the license and includes backend metadata for vLLM, SGLang and LMDeploy TurboMind.

Sources

  1. Qwen3-Coder-480B-A35B-Instruct model card
  2. Qwen3-Coder GitHub repository
  3. Qwen3-Coder-480B-A35B-Instruct on ModelScope

From Shadowfetch