DeepSeek releases DeepSeek Coder

DeepSeek, then a little-known Chinese lab spun out of the quantitative hedge fund High-Flyer, released DeepSeek Coder, a family of open-weight code models ranging from 1 billion to 33 billion parameters — its first public model release, preceding the general-purpose chat models that would later make the company widely known.

The models were trained from scratch on 2 trillion tokens, with a training mix DeepSeek said was 87% code and 13% natural language in English and Chinese, and used a project-level corpus intended to let the models reason about multi-file codebases rather than isolated snippets. DeepSeek reported that its largest base model, at 33 billion parameters, beat Meta’s CodeLlama-34B by several points across benchmarks including HumanEval and MBPP, and that an instruction-tuned version of the same model outperformed OpenAI’s GPT-3.5-turbo on HumanEval while matching it on MBPP — a claim of parity with a proprietary model from an open release, on code tasks specifically, that was unusual for the time.

The weights were released under a permissive licence allowing unrestricted commercial use, distinguishing the release from more heavily restricted open weights such as Llama 2’s. A fuller technical paper describing the training methodology followed in January 2024, after the models themselves had already been available and in use for several months.

DeepSeek Coder attracted comparatively little Western attention on release — DeepSeek was not yet a name most people outside specialist AI circles recognised — but it marked the start of a pattern the company repeated with its subsequent releases: shipping strong open-weight models with permissive licensing well ahead of the broader recognition that would follow once DeepSeek’s general chat and reasoning models drew wider notice in 2024 and 2025.