DeepSeek releases DeepSeek-Coder-V2
The 236B-parameter mixture-of-experts model scored 90.2% on HumanEval, edging out GPT-4-Turbo's 88.2%, while running with only 21B parameters active per token.
- Open weights & ecosystem
- Models & capabilities
- Minor
DeepSeek released DeepSeek-Coder-V2, an open-weight mixture-of-experts model for code and mathematics, following its original DeepSeek-Coder from late 2023. It shipped in two sizes: a 16B-parameter version with 2.4B active parameters, and a 236B-parameter version with 21B active, both supporting a 128,000-token context window and trained on an additional 6 trillion tokens beyond the base DeepSeek-V2 model.
DeepSeek reported that the larger Instruct variant scored 90.2% on HumanEval and 75.7% on the MATH benchmark, against 88.2% and 73.4% respectively for GPT-4-Turbo’s April 2024 version — figures the company presented as rough parity with, or a slight edge over, the leading closed model on code-specific tasks. Programming language coverage was also expanded, from 86 languages supported by the original DeepSeek-Coder to 338.
The code was released under the MIT licence, with model weights under a separate DeepSeek licence permitting commercial use. As with other DeepSeek releases, the benchmark comparisons came from the company itself rather than an independent evaluation, but the release reinforced a pattern through 2024 in which Chinese open-weight labs published coding and reasoning models whose reported scores matched or exceeded the current closed-source frontier, at a fraction of the parameter count actually activated per query.