Alibaba's Qwen2.5-Coder goes full open source, matching GPT-4o

Qwen2.5-Coder is a dedicated open-source code model family developed by the Qwen team. This release completes the lineup by adding 0.5B, 3B, 14B, and 32B variants to the existing 1.5B and 7B models, giving developers six different sizes across Base and Instruct checkpoints.

The Base models are tuned as clean foundations for custom pre-training or internal company fine-tuning. The Instruct models are fully aligned out of the box for direct chat, coding assistants, and agent setups.

What it can do

The star of the release is Qwen2.5-Coder-32B-Instruct. Across mainstream code generation benchmarks like EvalPlus, LiveCodeBench, and BigCodeBench, it sets a new open-source standard and matches GPT-4o in head-to-head performance.

Bug fixing is where it really shows its teeth. On the rigorous Aider benchmark for code repair, the 32B model scored 73.7, performing neck-and-neck with GPT-4o.

Language versatility is another highlight. It covers more than 40 programming languages, scoring 65.9 on McEval with notable fluency in niche languages like Haskell and Racket. It also took the top open-source spot on MdEval with a 75.2 score for multilingual code repair.

For real IDE workflows, it natively supports Fill-in-the-Middle (FIM) completions. Constrained to an 8K sequence length, it swept first place across five in-filling benchmarks, including RepoEval and CrossCodeEval.

Specs, licenses, and getting started

Context windows scale with model size. The lightweight 0.5B, 1.5B, and 3B models handle 32K context, while the larger 7B, 14B, and 32B versions support up to 128K context for scanning entire repositories.

The licensing terms are surprisingly permissive. Five of the sizes (0.5B, 1.5B, 7B, 14B, and 32B) are published under Apache 2.0, permitting commercial use. Only the 3B variant is gated behind a proprietary Qwen Research license.

You can pull the weights right now from GitHub, Hugging Face, ModelScope, and Kaggle.

To put it to work immediately, developers can configure it as a custom backend in Cursor for code completion, or plug it into Open WebUI to generate dynamic artifacts like mini-games, landing pages, and interactive data charts.

Why it matters

Until now, local coding models felt like second-tier backups next to closed frontier APIs like GPT-4o. Qwen2.5-Coder-32B changes that math by bringing flagship-tier coding and debugging to hardware you can actually control.

If you work with proprietary corporate codebases or simply want to stop paying recurring IDE subscriptions, this 32B release is good enough to become your primary coding engine today.