The original DeepSeek Coder supported 86 programming languages and a context window of 16K, but the new V2 version expands language support to 338 and context window to 128K. When tested on various benchmarks designed to evaluate code generation, editing, and problem-solving capabilities, DeepSeek Coder V2 scored higher than most closed and open-source models. The model also delivers decent performance in general reasoning and language understanding tasks. It is being offered under an MIT license, which allows for both research and unrestricted commercial use.
Key takeaways:
- Chinese AI startup DeepSeek has released DeepSeek Coder V2, an open-source mixture of experts (MoE) code language model that excels at both coding and math tasks and supports more than 300 programming languages.
- DeepSeek Coder V2 outperforms state-of-the-art closed-source models, including GPT-4 Turbo, Claude 3 Opus and Gemini 1.5 Pro, marking the first time an open model has achieved this.
- The model was trained on an additional dataset of 6 trillion tokens, largely comprising code and math-related data sourced from GitHub and CommonCrawl, enabling it to optimize for diverse computing and application needs.
- DeepSeek Coder V2 is being offered under a MIT license, allowing for both research and unrestricted commercial use, and can be accessed via Hugging Face or through the company's platform via API.