Qwen3-Coder 480B A35B: The Open Coder That Started the Line

Qwen3-Coder-480B-A35B-Instruct is the model most people mean when they say "Qwen3-Coder". Alibaba's Qwen team released it in July 2025 as the first member of a coding line that has since grown a hosted Plus tier, a cheaper Flash tier and, in February 2026, a much smaller successor called Qwen3-Coder-Next.

That makes it an awkward model to evaluate fifteen months later. It is no longer the newest thing in its own family, but it is still the largest open-weight model in the Qwen3-Coder line, it is still hosted by several providers, and its per-token prices are low. The useful question is not whether it is good, but whether it is still the right Qwen coder for your job.

What Qwen shipped in July 2025

The model is a mixture-of-experts design with 480B total parameters and 35B active per token. The model card lists 62 layers, grouped-query attention with 96 query heads and 8 key-value heads, and 160 experts with 8 activated per token. Native context is 262,144 tokens, and Qwen says it can be extended to 1M with YaRN extrapolation.

The weights are on Hugging Face under Apache 2.0, which is about as permissive as model licences get: you can fine-tune, redistribute and serve it commercially without a separate agreement.

It is a non-thinking model. It does not emit <think> blocks, and the card notes that passing enable_thinking=False is no longer needed. Every token you pay for is answer or tool call, not hidden deliberation.

How it was trained, and why that shapes behaviour

Qwen's launch post gives three details worth knowing. Pre-training used 7.5T tokens with a 70% code ratio, and the team used the earlier Qwen2.5-Coder to clean and rewrite noisy data. That is a heavily code-weighted diet, and it shows: the model is less of a general assistant than the general Qwen3 releases.

Post-training leaned on reinforcement learning in two forms. Code RL scaled test-case generation across a broad set of real coding tasks. Agent RL trained long-horizon, multi-turn tool use across what Qwen describes as 20,000 independent environments running in parallel on Alibaba Cloud.

The practical consequence is that this model was built to sit inside an agent harness — read files, run commands, read the output, try again — rather than to produce a perfect answer in one turn. Qwen open-sourced its Qwen Code CLI alongside it, adapted from Gemini CLI, and the model also works in Cline and other tools that speak the OpenAI protocol.

Settings that matter when you run it

The model card recommends temperature=0.7, top_p=0.8, top_k=20 and repetition_penalty=1.05, with an output budget of 65,536 tokens for most queries. If your harness pins a lower temperature for code, try the recommended values before concluding the model is weaker than its reputation.

Tool calling uses a model-specific function-call format, and Qwen's repository notes that SGLang and vLLM needed a new tool parser for it. If tool calls arrive as plain text instead of structured calls, the serving stack is the first suspect, not the model.

The model also supports fill-in-the-middle with the <|fim_prefix|>, <|fim_suffix|> and <|fim_middle|> tokens, so it can back an editor completion feature as well as an agent.

Self-hosting a 480B model is a cluster decision

The 35B active figure sets compute per token, which is why hosted prices are low. The 480B total figure sets memory, and that is what decides whether you can run it yourself. At BF16 the weights alone are far beyond a single GPU, and even an FP8 build, which Qwen publishes, needs a multi-GPU node before you account for KV cache at long context.

For most teams that settles it: rent it. OpenRouter lists it at $0.30 per million input tokens and $1.00 per million output at the time of writing, with cached input at $0.10. Active vs total parameters explains why those two numbers lead to such different answers on cost and on hardware.

Qwen3-Coder or Qwen3-Coder-Next

This is the comparison that matters in 2026. Qwen3-Coder-Next activates about 3B parameters out of 80B, uses a hybrid attention design, and ships under the same Apache 2.0 licence. On the figures each model card currently lists, Next scores higher on both agentic benchmarks they share:

• SWE-bench Pro — 38.7 for the 480B model, 44.3 for Next
• Terminal-Bench 2.0 — 23.9 for the 480B model, 36.2 for Next

Treat those as vendor numbers on vendor scaffolds, not as a controlled head-to-head. But the direction is clear, and Next does it with roughly a tenth of the active parameters. The detail is in the Qwen3-Coder-Next guide.

The 480B model still has a case. It has more total capacity for knowledge-heavy work — obscure libraries, older languages, unusual build systems — where a small active footprint can run out of recall. It has fifteen months of community tooling and quantisations behind it. And if your prompts already work well on it, a migration has a cost that a benchmark gap does not pay for automatically.

A short decision rule

• Pick the 480B model if you rent inference, your work is knowledge-heavy, and you have prompts tuned for it already.
• Pick Next if you want to self-host, care about throughput, or are starting fresh.
• Pick the hosted Plus tier if you want Alibaba to serve it with a 1M window; see the Qwen3-Coder-Plus guide.

Whichever way you lean, run both on ten tasks from your own repository before committing. How to benchmark LLMs on your own work has a harness you can set up in an afternoon.