Qwen3-Coder-480B-A35B: Repository-Scale Context, No Thinking Blocks

What Qwen3-Coder-480B-A35B-Instruct is

Qwen3-Coder is Alibaba's agentic code model line, and this is its largest variant: 480 billion total parameters with 35 billion activated per token, open-weight and self-hostable. Alibaba's claim for it is specific — significant performance among open models on agentic coding and agentic browser use, with results they describe as comparable to Claude Sonnet. The design target is repository-scale understanding rather than snippet completion, and the context window is sized accordingly.

256K natively, a million with YaRN

Native context is 262,144 tokens, and Alibaba state it is extendable to 1 million using YaRN. The extension is a configuration step rather than a default.

Non-thinking only

Alibaba are explicit that this model supports only non-thinking mode and does not generate thinking blocks in its output, adding that specifying enable_thinking=False is no longer required.

The architecture and the harnesses

The model is a causal language model of 62 layers with grouped-query attention at 96 query heads to 8 key-value heads, and a mixture-of-experts layer of 160 experts with 8 activated per token. That 8-of-160 sparsity is what keeps 480 billion parameters to a 35 billion active budget. On the tooling side Alibaba name Qwen Code and CLINE as supported platforms and note a specially designed function call format.

LMArena

On LMArena's text leaderboard Qwen3 Coder 480B A35b Instruct has 1388 Elo (rank 184); by category, 1457 Elo in Coding (rank 149), 1376 Elo in Math (rank 196), 1380 Elo in Expert (rank 204) and 1362 Elo in Creative Writing (rank 168).

Price

Roughly $0.40 per million input tokens through third-party providers.

Recommended settings

Alibaba recommend temperature 0.7, top_p 0.8, top_k 20 and a repetition penalty of 1.05, and an output length of 65,536 tokens for most queries. If memory runs out, they suggest reducing the context to 32,768 tokens. The model needs Transformers 4.51.0 or later, and Ollama, LM Studio, MLX-LM, llama.cpp and KTransformers support Qwen3 for local use. Tools are defined in the standard OpenAI function format and called through an OpenAI-compatible endpoint. Alibaba describe Qwen3-Coder as available in multiple sizes, with the 480B-A35B variant released first, and point to the Qwen3-Coder blog, GitHub repository and the Qwen3 technical report (arXiv 2505.09388) for benchmark results and hardware requirements.

Training and release

Alibaba released Qwen3-Coder on 22 July 2025. Pre-training used 7.5 trillion tokens with a 70% code ratio, while preserving general and mathematical abilities. Post-training applied code reinforcement learning to a broad set of real-world coding tasks with automatically scaled test cases, and long-horizon agent reinforcement learning on a system that runs 20,000 independent environments in parallel on Alibaba Cloud. Alibaba's launch post describes performance on agentic coding, agentic browser use and agentic tool use as comparable to Claude Sonnet 4, and states SWE-Bench Verified results without test-time scaling.

Qwen Code

Alongside the model Alibaba released Qwen Code, a command-line coding agent forked from Gemini CLI with an adapted parser and tool support for Qwen3-Coder. It installs through npm, requires Node.js 20 or later, and calls models through the OpenAI SDK.

Official Resources