Qwen3-Coder

2025-07-22 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba Qwen releases Qwen3-Coder-480B-A35B-Instruct, a MoE with 480 billion parameters (35 billion active) under Apache 2.0, dedicated to agentic coding, with the Qwen Code command-line tool.

### Context In mid-2025, coding agents (Claude Code, Gemini CLI) rely on models able to chain many tool calls. The vendor compares Qwen3-Coder to Claude Sonnet 4 and presents it as the "most agentic" of its coding models. ### What the model brings - Pre-training on 7.5 trillion tokens, 70% of which is code, with rewriting of noisy data by Qwen2.5-Coder. - Native context of 256,000 tokens, extensible to 1 million with YaRN, designed for entire repositories and pull requests. - Reinforcement learning on code ("hard to solve, easy to verify") followed by long-horizon agentic reinforcement, with 20,000 independent environments in parallel on Alibaba Cloud's infrastructure. - Results announced as state of the art among open models on SWE-Bench Verified without test-time scaling, and comparable to Claude Sonnet 4 according to the vendor. - Qwen Code tool, derived from Gemini CLI, and compatibility with Claude Code, Cline via a dedicated function-calling format. ### Architecture and specifications MoE with 480 billion total parameters and 35 billion active, 62 layers, 160 experts with 8 active, GQA with 96 query heads for 8 key-value heads. Non-thinking mode only. Native context 262,144 tokens. Apache 2.0 license. Weights on Hugging Face and ModelScope. Smaller versions followed, including Qwen3-Coder-30B-A3B (July 31, 2025). Knowledge cutoff: not disclosed. ### Cost and access | Context tier (qwen3-coder-plus, Alibaba Cloud API, current pricing consulted) | Input | Output | |---|---|---| | 0 to 32 K tokens | $1 | $5 | | 32 to 128 K | $1.8 | $9 | | 128 to 256 K | $3 | $15 | | 256 K to 1 M | $6 | $60 | Price per million tokens, USD. Self-hosting hardware not disclosed by the vendor. ### Benchmarks The comparisons (SWE-Bench Verified, Terminal-Bench, agentic tasks) appear as image charts in the post and the model card; no numeric table in text was found. No value is therefore reproduced here. ### Versus closed models The vendor claims that Qwen3-Coder achieves results comparable to Claude Sonnet 4 on agentic coding, browser use and tool use, and sets a new state of the art among open models on SWE-Bench Verified. These claims rely on vendor charts, whose values could not be verified in text. On costs, the vendor's API is priced by context tiers, with output at 5 dollars per million tokens for small requests, with no sourced comparison with closed APIs here. The full 480-billion-parameter model requires a multi-GPU infrastructure, which limits self-hosting, but the Apache 2.0 license allows commercial use without restriction. ### Key takeaways Qwen3-Coder is the milestone where Qwen enters the field of coding agents with an open model and tool. The comparison figures remain to be verified in the vendor's charts.

Tags : LLM, Open Source, Alibaba, Qwen, MoE, Apache 2.0, Code