Skip to content
← Models
Open weightsAlibaba (Qwen)/10 trustVerified
Qwen3-Coder-30B logo

Qwen3-Coder-30B

Alibaba's 30B MoE coding model (3B active) — agentic coding with 256K context, Apache 2.0.

Visit official page →

About

Qwen3-Coder-30B-A3B-Instruct (July 2025) is the efficient member of Alibaba's Qwen3-Coder family: 30B total parameters with only 3B active per token (MoE), 256K native context (1M with extrapolation), and strong agentic coding — tool use, multi-file edits, and long-horizon software tasks.

Apache 2.0 licensed — commercial use allowed. With 386k Hugging Face downloads it is the most-used Qwen3-Coder checkpoint. The MoE design means inference costs sit near a 3B dense model while capacity approaches much larger ones — the same efficiency story as the flagship Qwen3-Coder-480B, distilled to runnable size. Pairs well with Qwen Code / Claude Code-style agentic harnesses. For the full-fat version see qwen3-coder (already in the catalog).

Specs

Context window

256,000 tokens native (1M with YaRN extrapolation)

Pricing

Free — open weights (Apache 2.0), self-hosted, commercial use allowed.

Free quota

Free — download and run the weights yourself (Apache 2.0).

API access

Self-hosted (Transformers, vLLM, Ollama); API via Alibaba Cloud Model Studio and aggregators.

Weights

Open weights

Released

July 23, 2025

Benchmarks

Provider-claimed (Qwen team): competitive with leading open coding models on SWE-bench-style agentic evaluations — treat as provider-claimed. Independent: strong community reception for agentic coding; no standardized third-party scores verified.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

codecoding-agentsagent-developmentdebuggingautomation

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier). Apache 2.0, the most-used Qwen3-Coder checkpoint (386k downloads), and the MoE efficiency story is technically sound with strong community validation for agentic coding. Docked two points: (1) benchmark claims are provider-reported on agentic evals without independent verification; (2) the 256K→1M context extrapolation is provider-described — verify long-context quality for your use case.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.3/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
9.4/10
Manual review · sanity read + red-flag check
9.0/10
  • · HF downloads 386,097 (peer-normalized)
  • · released 447 days ago
  • · manual review: most-used Qwen3-Coder checkpoint; MoE efficiency sound; agentic coding validated by community; benchmarks provider-reported; license: Apache-2.0

More in the qwen3-coder family

Sources