Qwen3 Coder 480B A35B Instruct

This model utilizes a Mixture-of-Experts architecture, featuring 480 billion total parameters with 35 billion active parameters per forward pass, distributed across 160 experts. Designed specifically for agentic coding, it excels at multi-step workflows, function calling, and tool use. By supporting deep repository-scale reasoning, it is built to assist developers in creating functional applications, positioning itself as a high-performance alternative to proprietary models in tasks like browser-use and foundational coding.

Developed through comprehensive pre-training and post-training stages, the model incorporates specialized function call protocols to enhance its utility in developer environments. It features 62 layers and a GQA attention mechanism, ensuring efficient performance during complex reasoning. With its ability to handle extensive token sequences, the model is optimized for long-context understanding, making it a robust choice for developers seeking to integrate advanced, agent-driven coding capabilities into their existing software workflows.

submodelQwen/Qwen3-Coder-480B-A35B-Instruct-FP8qwen

Open provider page Provider docs

Quick Info

Provider: submodel
Model key: Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8
Release date: Aug 23, 2025
Last updated: Aug 23, 2025
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.20
Output token cost: $0.80

Limits

Output tokens: 262,144 tokens
Context window: 262,144 tokens

Transparent token rates

Compare Qwen3 Coder 480B A35B Instruct pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Qwen3 Coder 480B A35B Instruct

submodel

CoverageBenchmark

Kilo Code's model directory entry restates the OpenRouter pricing tier for Qwen3 Coder 480B A35B at $0.97 per 1M input and $4.88 per 1M output tokens (matching the Alibaba OpenSource listing), with a 262,144-token context and 65,536-token max output. It frames the model inside the Kilo Code coding agent, which integrat Independent benchmark figures sourced from Artificial Analysis give a coding-focused performance snapshot: 58.5% on LiveCodeBench for real-world coding tasks, 35.9% on SciCode for scientific computing, 18.9% on TerminalBench Hard for CLI generation, 45.3% on LCR for long-context reasoning, and 40.5% on IFBench for inst

submodel

CoverageBenchmark

OpenRouter lists Qwen3 Coder 480B A35B as a 480B-total / 35B-active Mixture-of-Experts code model from the Qwen team, with a 262K context window, July 22 2025 release, and a June 2025 knowledge cutoff. The page is optimized for agentic coding tasks including function calling, tool use, and long-context repository reaso The OpenRouter routing index shows five live providers for the same weights: Google Vertex ($0.22/$1.80, ~51 t/s, 0.74s latency), DeepInfra Turbo ($0.30/$1.00, 37 t/s, 0.42s), Venice ($0.35/$1.50, 17 t/s), Alibaba OpenSource ($0.975/$4.875, 7 t/s), and NovitaAI ($0.38/$1.55, 32 t/s), each reporting its own cache-hit an