Qwen/Qwen3-Coder-30B-A3B-Instruct

Built on the Qwen3 architecture, this model utilizes a sparse Mixture-of-Experts design to balance computational efficiency with high-level reasoning. With a total of 30.5 billion parameters and 3.3 billion active parameters per forward pass, it is engineered to handle complex programming tasks and repository-scale understanding. The architecture features 48 layers and 128 experts, allowing it to maintain performance while remaining streamlined. It is specifically optimized for agentic coding and browser-use tasks, providing a robust foundation for developers who require precise, structured code generation and reliable tool-use capabilities.

The model underwent extensive pre-training and post-training to refine its instruction-following abilities, specifically for environments that require direct, non-thinking responses. By supporting a native context length of 262,144 tokens—which can be extended to 1 million tokens using Yarn—it is well-suited for analyzing massive codebases and long-form documentation. Its design lineage emphasizes seamless integration with standard tool-use formats, making it a practical choice for automated coding assistants and complex agentic workflows that demand both depth of context and high-speed inference.

SiliconFlow (China)Qwen/Qwen3-Coder-30B-A3B-Instructqwen

Open provider page Provider docs

Quick Info

Provider: SiliconFlow (China)
Model key: Qwen/Qwen3-Coder-30B-A3B-Instruct
Release date: Aug 1, 2025
Last updated: Nov 25, 2025
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.07
Output token cost: $0.28

Limits

Output tokens: 262,000 tokens
Context window: 262,000 tokens

Transparent token rates

Compare Qwen/Qwen3-Coder-30B-A3B-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 Qwen/Qwen3-Coder-30B-A3B-Instruct

SiliconFlow

CoverageBenchmark

The OpenRouter listing for Qwen3-Coder-30B-A3B-Instruct describes it as a 30.5B-parameter Mixture-of-Experts model with 128 experts and 8 active per forward pass, built on the Qwen3 architecture and aimed at advanced code generation, repository-scale understanding, and agentic tool use. It supports a native 256K-token Benchmark figures aggregated by OpenRouter from Artificial Analysis and Design Arena include GPQA Diamond 51.6%, HLE 3.8%, IFBench 32.7%, τ²-Bench Telecom 34.5%, AA-LCR 32.7%, CritPt 0.0%, Terminal-Bench Hard 15.2%, AA-Omniscience Accuracy 16.4%, and AA-Omniscience Non-Hallucination Rate 19.7%, alongside Design Arena E

SiliconFlow

Coverage

The LLM Explorer catalog page for Qwen3 Coder 30B A3B Instruct lists it as an open-source Apache-2.0 language model from Qwen with 30B parameters, a MoE architecture (Qwen3MoeForCausalLM), 256K-token context, instruction-following and code-generation capabilities, and links to the Hugging Face repo at huggingface.co/Qw The page also surfaces community alternatives and sibling Qwen3 30B A3B variants (including YOYO V2/V3 finetunes and a REAP 25B A3B pruning), as well as links to the underlying Arxiv paper 2505.09388 and the model card, framing Qwen3-Coder-30B-A3B-Instruct within the broader Qwen3 MoE family. The listed "Updated 2026-0