DeepSeek-V3.2 Arrives With FP8 Weights, MIT License
The new Mixture-of-Experts model from DeepSeek AI combines an efficient FP8 architecture with a fully permissive license for commercial use.
DeepSeek AI has released DeepSeek-V3.2, a new open-source language model that stands out for its technical efficiency and permissive licensing. The model uses a Mixture-of-Experts (MoE) architecture, placing it among a class of powerful yet computationally lean models designed for scalable inference.
The most notable technical feature is its use of FP8 (8-bit floating point) weights. This precision format significantly reduces the model's memory footprint compared to traditional 16-bit models. For developers, this translates to faster inference and the ability to run the model on more accessible, consumer-grade hardware.
The MoE design complements this focus on efficiency. By routing input tokens to specialized "expert" sub-networks, the model can achieve the performance of a much larger dense model while only activating a fraction of its parameters for any given task, further reducing computational costs.
Perhaps most importantly for the open-source community, DeepSeek-V3.2 is available under the MIT License. This choice removes many of the commercial use restrictions seen with other high-performance models, offering developers and businesses a truly open foundation for building new applications.
Sources
- Visit
deepseek-ai/DeepSeek-V3.2
Hugging Face
More from DeepSeek
All DeepSeek releases →
DeepSeek V4.1-Flash Adds Vision and Reasoning
The new MIT-licensed checkpoint folds image understanding and reasoning into DeepSeek's fast, mixture-of-experts Flash line.
DeepSeek adds vision to its V4 Flash line
An experimental, MIT-licensed vision-language model brings image understanding to DeepSeek's fast V4 Flash architecture.
DeepSeek Releases V4-Pro-0813 With Open Weights
The Chinese lab pushes a higher-capability checkpoint of its V4 line to Hugging Face under a permissive MIT license.
More in Text / LLM
All Text / LLM →IBM's Granite Switch Treats Models Like Software
A new Apache-2.0 release reframes enterprise LLMs as components you can swap and update in place.
Mistral Large 4 Arrives as a 1T-Parameter MoE
The new flagship pairs a trillion total parameters with 52B active per token and adds vision alongside multilingual text.
Nanbeige 4.3 Arrives as a 33B Bilingual Model
A dense English–Chinese language model lands under Apache 2.0, aimed at teams that want permissive licensing at mid-size scale.
0 comments
No comments yet. Be the first to weigh in.