PFN releases PLaMo 3.0 Fin, an 8B finance-focused model: Japanese financial benchmark rises from 48.0 to 55.8

Financial institutions must process large volumes of text quickly, and routing all of it through large models raises API and compute costs and slows throughput, while customer and transaction data may require closed cloud or on-premises deployment; PLaMo 3.0 Fin is released at 8B scale for such settings, with weights on Hugging Face under the PLaMo Community License. The base model, PLaMo 3 NICT 8B Base, comes from joint research by PFN and NICT and carries no finance-specific continued training. On the five tasks of japanese-lm-fin-harness, the base model averages 48.0, continued pretraining reaches 50.1, supervised fine-tuning reaches 56.0, and direct preference optimization gives 55.8.

The model writes a Japanese reasoning trace before answering and can call external tools such as web search, calculation and database lookup, so a user can check which formula and numbers produced an answer. Training applies continued pretraining on financial corpora, then supervised fine-tuning, then direct preference optimization to the base model, and the fine-tuning stage learns instruction following, Japanese reasoning traces and tool calls together. The demo environment holds about 90 tools, but the model initially receives only search_tools and calls the tools it retrieves; the fine-tuning data was built with Gemma 4 31B generating fictitious tools and step-by-step tool search, calls and answers.

The demonstrated scope is Japanese financial knowledge questions, finance-themed multi-turn dialogue and PFN's own Japanese financial tool-use benchmark; the tool-use evaluation harness and data are not public, and comparator values come from measurement in the same environment. The authors recommend use in settings with few tool candidates, or narrowing the tools passed at once through a retrieval step as in the demo.

With many tool candidates the model can fail to settle on one, extending its reasoning until the context limit and producing no answer, as happened in 5 to 10 of 540 tool-use benchmark samples; at 8B scale its stored knowledge is limited, and it may still give plausible wrong answers on unfamiliar topics, so the authors recommend retrieving evidence with tools and having a human check answers where facts and numbers matter.