Model

Qwen3-Coder-480B-A35B

OPEN

Alibaba · Qwen3-Coder family · released Jul 22, 2025

Most powerful open agentic coding MoE; 256K native (1M extrapolated) context, rivals Claude Sonnet 4.

ReasoningCodingVisionFunction callingTool useAgentic

1364.5

Elo · rank #145

Parameters

480B

Active params

35B (MoE)

Context

256K tokens

Architecture

MoE, 160 experts / 8 active

License

Apache 2.0

Languages

119+

API price (in/out)

$1.5 / $7.5

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

Aider PolyglotCoding61.8%#17

best: Claude Opus 4.5 · 89.4%

BFCL v3Agents68.7%#12

best: Hunyuan-A13B · 78.3%

GPQA DiamondReasoning62.0%#159

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning4.0%#134

best: Claude Sonnet 5 · 57.4%

LiveCodeBenchCoding44.9%#125

best: DeepSeek-V4-Pro (Think Max) · 93.5%

SWE-bench ProCoding38.7%#41

best: Claude Fable 5 · 80.0%

SWE-bench VerifiedCoding66.5%#66

best: Claude Fable 5 · 95.0%

tau-benchAgents60.0%#15

best: Claude Opus 4.1 · 82.4%

Terminal-Bench 2.0Coding23.9%#56

best: GPT-5.6 · 88.8%

Run it locally

VRAM @ Q4

289 GB

VRAM @ FP16

960 GB

Fits on (Q4)

M3 Ultra 512GB

Far too large for a single RTX 4090; needs multi-GPU or large CPU-RAM offload setups.

Quantizations

GGUF · FP8 · AWQ

Qwen3-Coder family

Elo progression across releases

Qwen3-Coder-480B-A35BJul 20251364.5

Resources

HFQwen/Qwen3-Coder-480B-A35B-Instruct GHQwenLM/Qwen3-Coder DOCSqwen.readthedocs.io APIAvailable via Alibaba API

API price $1.5/$7.5 · each benchmark row carries its own source badge (see methodology)