Qwen3's largest coding-specialized model (480B total, 35B active, 160 experts, 8 active per token MoE). State-of-the-art code generation and understanding. Non-thinking mode only.
Qwen3-Coder-480B-A35B-Instruct is Alibaba Qwen's flagship open-weight coding model: a 480B-total / 35B-active Mixture-of-Experts model built for agentic coding, repository-scale understanding, and tool use. Qwen describes it as its most agentic code model to date, with state-of-the-art open-model results on agentic coding, browser-use, and tool-use tasks and performance positioned against Claude Sonnet-class systems.
For TheRouter users, this is a practical default candidate for AI coding teams that want a large open-weight coder behind Cursor, Cline, Claude Code-style routers, or any OpenAI-compatible client. It exposes 256K native context, can be extended to 1M tokens with YaRN in self-hosted setups, and is explicitly non-thinking: it does not emit <think> blocks, so production clients can treat its responses as direct coding output.
| Type | Rate |
|---|---|
| Input | $1.62 / 1M tokens |
| Output | $8.10 / 1M tokens |
| Architecture | Mixture-of-Experts causal language modelhuggingface.co ↗ | verified |
| Parameters | 480B total / 35B activatedhuggingface.co ↗ | verified |
| Experts | 160 experts; 8 activated per tokenhuggingface.co ↗ | verified |
| Native context | 262,144 tokens; extendable to 1M with YaRN in self-hosted deploymentsqwenlm.github.io ↗ | verified |
| Output mode | Non-thinking only; does not generate <think> blockshuggingface.co ↗ | verified |
| Training emphasis | 7.5T pretraining tokens, about 70% code; code RL and long-horizon agent RLqwenlm.github.io ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Verified Reported by Together AI for Qwen3-Coder-480B-A35B-Instruct; use as a secondary-source benchmark until an official numeric table is reachable. | 69.6%% | together.ai ↗ | |
Aider Polyglot Multilingual code-editing benchmark; secondary public report, not an official Qwen table in the fetched sources. | 61.8%% | medium.com ↗ | |
Agentic Coding / Browser-Use / Tool-Use Official qualitative claim from Qwen announcement; no exact score exposed in the fetched page text. | State-of-the-art among open models | qwenlm.github.io ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "qwen/qwen3-coder-480b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Use the standard OpenAI-compatible chat completions endpoint. This model is strongest when the prompt contains concrete files, failing tests, or explicit coding constraints.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/qwen3-coder-480b",
"messages": [
{"role": "system", "content": "You are a senior software engineer. Return concise patches and explain tradeoffs."},
{"role": "user", "content": "Refactor this TypeScript function to remove duplication and preserve behavior: ..."}
],
"temperature": 0.7,
"max_tokens": 4096
}'The Qwen3-Coder-Next report frames 480B-A35B as the previous flagship coder and shows Qwen continuing to optimize coding-agent performance per active parameter, not just total model size.
Qwen introduced its flagship 480B/35B-active coding MoE alongside Qwen Code, a Gemini Code-derived CLI adapted with Qwen-specific prompts and function-calling protocols.
The model card documents 62 layers, 96 Q heads / 8 KV heads, 160 experts with 8 active, 262,144-token native context, and the requirement that the model only runs in non-thinking mode.
No. Qwen explicitly states that this checkpoint supports only non-thinking mode and does not generate <think></think> blocks. That is useful for coding clients because responses are direct, but it also means you should not build UX around visible reasoning traces.
The native context length is 262,144 tokens. Qwen's launch post says the model can be extended up to 1M tokens with YaRN, but that is a deployment/configuration path for self-hosters rather than a guarantee that every hosted route exposes 1M context.
Yes, if the client can speak OpenAI-compatible chat completions or route through a compatible proxy. The Qwen launch post specifically discusses Qwen Code, Cline, and Claude Code-style integration paths; through TheRouter, the simplest path is the OpenAI-compatible /v1/chat/completions endpoint with model qwen/qwen3-coder-480b.
480B is the large flagship: 480B total parameters and 35B active parameters. The 30B sibling is much cheaper and faster, with 30B total / 3B active parameters. Use 480B for difficult multi-step repository work; use 30B for cheaper autocomplete, small edits, and high-throughput IDE assistance.
Yes, the weights are published and Qwen provides BF16 and FP8 checkpoints. The practical constraint is hardware and serving-stack support: use recent SGLang or vLLM with the qwen3_coder tool-call parser, and reduce context length if memory pressure prevents a full 256K deployment.
| source | URL | retrieved | |
|---|---|---|---|
| Architecture | huggingface.co ↗ | 2026-06-02 | verified |
| Parameters | huggingface.co ↗ | 2026-06-02 | verified |
| Experts | huggingface.co ↗ | 2026-06-02 | verified |
| Native context | qwenlm.github.io ↗ | 2026-06-02 | verified |
| Output mode | huggingface.co ↗ | 2026-06-02 | verified |
| Training emphasis | qwenlm.github.io ↗ | 2026-06-02 | verified |
| SWE-bench Verified | together.ai ↗ | 2026-06-02 | to verify |
| Aider Polyglot | medium.com ↗ | 2026-06-02 | single source |
| Agentic Coding / Browser-Use / Tool-Use | qwenlm.github.io ↗ | 2026-06-02 | verified |
| Qwen3-Coder-Next technical report compares the new 80B/3B model to Qwen3-Coder-480B | arxiv.org ↗ | 2026-06-02 | verified |
| Qwen launches Qwen3-Coder-480B-A35B-Instruct and Qwen Code | qwenlm.github.io ↗ | 2026-06-02 | verified |
| Hugging Face card publishes architecture and non-thinking-mode details | huggingface.co/Qwen ↗ | 2026-06-02 | verified |
| Is Qwen3-Coder-480B a thinking model? | huggingface.co ↗ | 2026-06-02 | to verify |
| How much context does it support? | qwenlm.github.io ↗ | 2026-06-02 | to verify |
| Can I use it from Cursor, Cline, or Claude Code-style tools? | qwenlm.github.io ↗ | 2026-06-02 | to verify |
| What makes it different from qwen/qwen3-coder-30b? | github.com ↗ | 2026-06-02 | to verify |
| Can I self-host Qwen3-Coder-480B? | huggingface.co ↗ | 2026-06-02 | to verify |