Qwen2.5-Coder models

Strongest: Qwen2.5-Coder-32B-Instruct

Alibaba has since superseded it with the Qwen3-Coder line — Qwen3-Coder-480B-A35B (2025-07-22), Qwen3-Coder-30B-A3B (2025-07-31) and Qwen3-Coder-Next (2026-01-30), none of them in this catalog — and Model Studio no longer serves the Qwen2.5 series; the open weights remain on Hugging Face. Qwen2.5-Coder-32B-Instruct is a code-specialist fine-tuned from Qwen2.5-32B-Base on a code-heavy mix, shipped 2024-11-12 under Apache 2.0. It was the first open-weight model to beat GPT-4o on HumanEval (92.7 vs 90.2) and became the canonical local coding model — widely documented running on a 64GB MacBook Pro. The buyer's sentence: a self-hosted Copilot-grade code model, single-GPU, Apache-licensed, with first-class fill-in-the-middle for IDE autocomplete.