On Monday, the Chinese AI startup DeepSeek unveiled a new model, DeepSeek-V3-0324. Boasting an impressive 685 billion parameters, this model made a quiet debut on the AI repository Hugging Face without the fanfare of a press release, yet it has nonetheless sparked significant interest within the industry. As an enhanced iteration of DeepSeek-V3, the new model maintains the same number of parameters as its predecessor but introduces versatility with support for multiple precision floating-point formats in both inference and training calculations. Moreover, it is fully open-source, promising potential performance enhancements and bug resolutions.
