Osegi
AI·1 min read

PFN releases PLaMo 3.0 Fin, a small finance AI for one GPU

Preferred Networks has published PLaMo 3.0 Fin, a finance-focused small language model with open weights on Hugging Face that runs on a single GPU and allows commercial use.

Illustration of a compact AI language model running on a single GPU inside a secure banking server environment.
Image: Impress Watch
SUMODialoediviewergiftdetect

Preferred Networks (PFN) has released PLaMo 3.0 Fin, a small language model built for practical use at financial institutions. The company published it as an open-weight model on Hugging Face. It can be deployed in closed cloud environments or on-premises, and commercial use is permitted.

The model builds on PLaMo 3 NICT, a pretrained model developed in joint research with Japan's National Institute of Information and Communications Technology (NICT). PFN then trained it further on financial data centered on Japan. It includes a reasoning function that works through problems step by step and prints its reasoning in Japanese, which makes it easier to check the basis of an answer.

Built for high-volume banking tasks

PLaMo 3.0 Fin can call external tools such as web search, calculations and internal databases. Because it is a compact model, it can run from a single GPU. PFN envisions it being embedded in tasks such as classifying inquiries, extracting, converting and matching information from application documents, fraud detection and analyzing the business condition of borrowers.

The pitch addresses two pain points. Financial institutions handle large volumes of processing, and using a large language model for everything can drive up costs. They also work with highly confidential information, so efficient operation in closed environments matters.

Licensing and domestic technology

The model is offered under the PLaMo Community License. Commercial use, including internal operations and integration into a company's own products and services, is allowed within the scope the license defines.

PFN also used its own AI chip, MN-Core 2, for the instruction-following training behind the reasoning function and external tool use. The company says the whole stack, from the AI semiconductor to the foundation model and the finance-specific model, was developed with domestic Japanese technology.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:preferred networksplamofinance aiopen-weight modelhugging facemn-core 2

More from this day

See all →

More in: AI

See all →