Qwen3-Coder
Open agentic coding model with 256K native context (up to 1M via extrapolation) + tool-use
Qwen3-Coder is a top choice when you want an open(-weight) agentic coding model with very long context for repo-scale work and a practical tooling story (Qwen Code CLI, OpenAI-compatible endpoints). It’s especially attractive for teams building custom coding agents, but you should validate latency/cost at your required context tier.
Why we love it
- Open model family with strong agentic coding focus and tool-use support
- 256K native context and 1M extrapolation targets repo-scale reasoning
- Clear integration paths via Qwen Code CLI and OpenAI-compatible APIs
Things to know
- Very large variants can be heavy to run; hosting/provider choice matters
- Tiered pricing by context length can make ultra-long prompts significantly more expensive
- Tool-calling format requires the updated tokenizer/tool parser for best results
About
Qwen3-Coder is Qwen’s open model family specialized for agentic programming: it’s designed for long-context repo understanding, tool use/function calling, and multi-step coding workflows. Flagship variant: Qwen3-Coder-480B-A35B-Instruct is a Mixture-of-Experts model (480B total parameters with 35B active) aimed at high-end coding and agent tasks. Long-context advantage: It supports 256K tokens natively and can be extended up to 1M tokens using extrapolation methods (e.g., YaRN), making it suitable for repo-scale reasoning and PR-sized dynamic inputs. Agent tooling ecosystem: Alongside the model, Qwen open-sources Qwen Code (a CLI adapted from Gemini CLI) and provides official paths to run Qwen3‑Coder through Claude Code and Cline via OpenAI-compatible endpoints. Practical coding features: The GitHub reference shows support for fill-in-the-middle (FIM) style code insertion and a dedicated tool parser for function calling. Pricing (Price-to-Value): If you use the hosted API via Alibaba Cloud Model Studio, qwen3-coder-plus is billed per input/output token with tiered pricing by context length; for 0–32K tokens the list price starts at $1/1M input tokens and $5/1M output tokens, increasing at higher context tiers.
Key Features
- ✓256K native context (extendable up to 1M with extrapolation) for repo-scale understanding
- ✓Agentic coding focus: tool use/function calling and long-horizon workflows
- ✓Open-source Qwen Code CLI for agentic development workflows
- ✓Works with popular agent interfaces (Claude Code, Cline) via OpenAI-compatible endpoints
- ✓Supports fill-in-the-middle (FIM) code insertion patterns
Frequently Asked Questions
Qwen3-Coder is Qwen’s open model family specialized for coding and agentic programming, emphasizing long-context repo understanding and tool use.
Qwen3-Coder supports 256K tokens natively and can be extended up to 1M tokens via extrapolation methods (e.g., YaRN).
Yes. Qwen’s official blog provides a setup using an API key from Alibaba Cloud Model Studio and configuring Claude Code to point to a DashScope proxy endpoint.