The updated DeepSeek-V3 'DeepSeek-V3-0324' is faster in all tests, and some say it has become the 'best non-inference model.'



Chinese AI company

DeepSeek has released an updated version of 'DeepSeek -V3-0324,' its version of 'DeepSeek -V3 ,' which was released in December 2024 as a model comparable to OpenAI's GPT-4o.

deepseek-ai/DeepSeek-V3-0324
https://simonwillison.net/2025/Mar/24/deepseek/



DeepSeek's New 641GB AI Model Lands Quietly — and Runs Surprisingly Fast on a Mac - WinBuzzer

https://winbuzzer.com/2025/03/24/deepseeks-new-641gb-ai-model-lands-quietly-and-runs-surprisingly-fast-on-a-mac-xcxwbn/

DeepSeek-V3 was announced in December 2024 as the largest large-scale language model ever, with 671 billion parameters.

Chinese AI company DeepSeek releases AI model 'DeepSeek-V3' comparable to GPT-4o, with a threatening 671 billion parameters - GIGAZINE



DeepSeek-V3 was a custom license, but the newly released DeepSeek-V3-0324 is released under the MIT license, an open source software license, and has a total file size of 641GB. Although it is not a new model, it has built-in FP8 quantization support. Also, although the number of parameters is 685 billion, only about 37 billion of them are active during inference, which relaxes the hardware requirements.

mlx-community/DeepSeek-V3-0324-4bit · Hugging Face
https://huggingface.co/mlx-community/DeepSeek-V3-0324-4bit

Machine learning researcher Auni Hanun ran a quantized version on a Mac Studio equipped with an Apple M3 Ultra chip and 512GB of unified memory and observed inference speeds of over 20 tokens per second.


Xeophon, an AI researcher, also ran a benchmark and reported that the score jumped out over DeepSeek-V3 in all tests, surpassing Anthropic's Claude 3.5 Sonnet and commenting that it was 'the best non-inference model.'



News site WinBuzzer notes that China has restrictions on chip imports from the US, making lightweight and efficient architectures like DeepSeek's valuable.

Related Posts:

in Software, Posted by logc_nt