Preferred Networks opens PLaMo 3.0 Fin finance SLM on Hugging Face

Preferred Networks developed PLaMo 3.0 Fin on top of its domestic PLaMo generative AI base model, targeting practical use inside financial institutions. The company published it as an open-weight model on its Hugging Face page on November 9, making the weights available for others to use.

The post-training for PLaMo 3.0 Fin included instruction-following for reasoning and tool use. This was carried out on MN-Core 2, an AI chip the company designed with a focus on matrix operations and performance per watt, using its PFCP computing platform. The company says a new program-porting method centered on a coding agent let developers without deep MN-Core 2 expertise port the instruction-following training code, including support for multi-board distributed training, in under a week. Training with the ported code achieved practical training speeds equivalent to a GPU-based environment, indicating MN-Core 2 can be a viable alternative to GPUs in AI development.