Qwen3 Coder 480B-A35B

Qwen3 Coder 480B A35B

Quick answer: Qwen3 Coder 480B A35B is Alibaba's July 2025 coding-specialised MoE model — 480B total parameters with 35B active at inference. It scores 7.7% on AetherCode. Under Apache 2.0 license with open weights available on Hugging Face. Designed for code generation, completion, and software engineering tasks.

At a Glance

Where Qwen3 Coder 480B A35B leads

• Apache 2.0 license — unrestricted commercial use
• Coding-specialised training — purpose-built for software engineering tasks
• 480B MoE with 35B active params — knowledge-rich but inference-efficient
• Open weights — self-hostable for high-volume code generation
• March 2025 knowledge cutoff (more recent than Qwen3 32B)

Where it lags

• 7.7% AetherCode — below Qwen3 32B (16.3%) on this specific benchmark
• Requires high-memory infrastructure to run the full 480B MoE
• 7.7% is significantly below proprietary models (o4-mini-high 35.5%)

Best for: Open-weight code generation at maximum parameter scale; organisations requiring Apache 2.0 licensing with a coding-specialised model; research into large MoE coding models.

What Qwen3 Coder 480B A35B Is

Qwen3 Coder 480B A35B is the large coding-specialised variant in Alibaba's Qwen3 Coder series, released July 1, 2025. It is a Mixture of Experts (MoE) model with 480B total parameters and 35B active at each forward pass — making it more inference-efficient than a comparable dense 480B model while retaining broad knowledge capacity.

The model is specifically trained for coding tasks: code completion, bug fixing, code explanation, and software engineering workflows. As an Apache 2.0 open-weight model, it is one of the most capable freely-licensable coding models available.

The 7.7% AetherCode score is lower than Qwen3 32B (16.3%) on this specific benchmark, which may reflect the particular evaluation set composition rather than general coding capability. Teams evaluating coding models should consider running multiple benchmarks relevant to their specific task type.

Specifications

Pricing

Qwen3 Coder 480B A35B is available as open weights under Apache 2.0 — free to self-host. Hosted API access available via Alibaba Cloud DashScope and third-party providers.

Public Benchmark Scores

Qwen3 Coder 480B A35B vs Alternatives

Qwen3 32B achieves a higher AetherCode score (16.3%) with far lower infrastructure requirements (32B dense vs 35B active from a 480B MoE). For AetherCode-type tasks, Qwen3 32B is the recommended open-weight choice unless the specific benefits of the 480B coder-specialised training are needed.

Related Models & Benchmarks

• Compare: Qwen3 32B · Kimi K2 Instruct · o4-mini
• Benchmarks: AetherCode

Frequently Asked Questions

Specs from Qwen's official Qwen3 Coder release (July 2025) and Benchgen evaluations. Last updated 2026-07-24.