Qwen3-Coder-30B-A3B
The lean MoE coder of the Qwen3-Coder line: 30B total / 3B active, 256K native context (1M with YaRN), Apache 2.0.
About
Qwen3-Coder-30B-A3B-Instruct is the smaller sibling of Alibaba's Qwen3-Coder family, released in late July 2025. It's a sparse mixture-of-experts model with 30.5 billion total parameters and only about 3.3 billion active per token (128 experts, 8 active), built specifically for agentic coding: tool use, function calling, and repository-scale understanding. Native context is 256K tokens, extendable to 1M with YaRN, and the weights are Apache 2.0. On Hugging Face it has pulled over 15 million all-time downloads, making it one of the most-run open coding models in existence.
In independent-style harness runs it resolved 51.6% of SWE-bench Verified (258/500) under OpenHands, and community testing put it at 60.9% on Aider Polyglot — roughly in the neighborhood of Claude Sonnet 4 and GPT-4.1 for agentic coding edits. It slots into Qwen Code, Cline, Roo Code, and most agent scaffolds, and runs on a single high-memory GPU or even Apple Silicon via MLX. The 480B big sibling scores higher on raw benchmarks, but for local agentic coding the 30B is the practical one.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- SWE-bench Verified 51.6% (258/500, OpenHands harness — SWE-bench community experiments repo, Aug 2025)
- Aider Polyglot 60.9% (community-reported third-party run; rivals Claude Sonnet 4 / GPT-4.1 per that report)
- Qwen's own blog positions it as top-tier among open models for agentic coding (provider framing, not a numeric claim)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: HF repo created Jul 31, 2025 with 15.6M all-time downloads and 1,290 likes; Apache 2.0 license on the model card; MoE config (128 experts, 8 active) confirmed in the model config; SWE-bench Verified 51.6% in the public SWE-bench experiments repo. Open weights. Release facts corroborated across the Qwen blog, HF, and community deployment reports. Strengths: elite adoption, genuinely permissive license, runs on modest hardware, deep agent-scaffold support. Docked for: (1) headline coding-benchmark numbers come from community harness runs, not Alibaba's own eval table; (2) Alibaba's published comparisons emphasize the 480B sibling. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 10.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · HF repo Qwen/Qwen3-Coder-30B-A3B-Instruct created 2025-07-31; 15.6M all-time downloads, 1,290 likes
- · MoE 30.5B total / ~3.3B active (128 experts, 8 active); 256K native context, 1M with YaRN
- · SWE-bench Verified 51.6% (258/500, OpenHands) in the public SWE-bench experiments repo
- · Apache 2.0 on the HF model card; supported by Qwen Code, Cline, Roo Code, vLLM, Ollama, MLX