Yi-Coder: A Compact Yet Powerful Code LLM Released by 01.ai

01.ai has introduced Yi-Coder, a series of small but highly capable open-source large language models (LLMs) tailored for code generation. Despite having fewer than 10 billion parameters, Yi-Coder models offer cutting-edge performance, available in 1.5B and 9B parameter variants.

The Yi-Coder-9B, the flagship model, has been pretrained on 2.4 trillion high-quality tokens sourced from GitHub and CommonCrawl, covering 52 programming languages. This model offers a unique 128K token context window, allowing it to handle project-level comprehension and code generation tasks with ease.

Yi-Coder-9B surpasses competitors in its parameter class, outperforming models such as CodeQwen1.5 and CodeGeex4. In competitive coding benchmarks like LiveCodeBench, Yi-Coder-9B-Chat achieved a 23.4% pass rate, surpassing even larger models such as DeepSeek-Coder-33B.

Beyond contest-level evaluations, Yi-Coder also excelled in code editing and completion tasks and became the first open-source code model to break 50% accuracy on CRUXEval-O. Its long-context comprehension and mathematical reasoning abilities make it a versatile tool for developers.

Yi-Coder models are now available as part of the open-source Yi family, and developers are encouraged to explore its features across various frameworks.