Alibaba Open-Sources Qwen2.5-Coder Series: SOTA Open-Source Code Model
Alibaba/Qwen
Alibaba Qwen has open-sourced the Qwen2.5-Coder series, including models from 0.5B to 32B parameters. The flagship 32B Instruct model achieves SOTA performance among open-source code models, matching GPT-4o on several benchmarks.
Alibaba Qwen announced the open-source release of the Qwen2.5-Coder series, claiming it is powerful, diverse, and practical. The 32B Instruct model achieves SOTA among open-source code models on benchmarks like EvalPlus, LiveCodeBench, BigCodeBench, and Aider, matching GPT-4o's coding capabilities. The series covers six sizes: 0.5B, 1.5B, 3B, 7B, 14B, and 32B, with Apache 2.0 license for most sizes except 3B which is under Qwen Research license. The models demonstrate strong code generation, repair, reasoning, and multi-language support. Qwen2.5-Coder also shows practical applicability in code assistants like Cursor and Artifacts scenarios. The team is working on code-centric reasoning models.
- Сокращения
- SOTA = State of the Art
- LLM = Large Language Model
- OOD = Out-of-Distribution
Source: Alibaba Qwen —
original