PFN Open-Sources PLaMo 3.0 Fin: an 8B Japanese Finance Model with Japanese Reasoning and Tool Calling, Scoring 55.8 on a Japanese Finance Benchmark Ahead of Gemma 4 E4B and gpt-oss-20b, Fine-Tuned on In-House MN-Core 2 Chips

Background and Industry Pain Points

Banks and brokerages handle huge volumes of inquiry triage, KYC document extraction and transaction monitoring alerts. Sending all of it to large-model APIs is costly and slow, and customer data often cannot leave the private network. PLaMo 3.0 Fin, released as open weights by Japan's Preferred Networks (PFN) on November 9, targets that gap: an 8B finance-specialized model that runs on a single GPU in a private cloud or on-premises, positioned as the core model for finance agents. It is PFN's second open release this month, after the 0.8B to 31B PLaMo 3 Translate models on November 5.

Architecture Highlights

The model starts from PLaMo 3 NICT 8B Base, co-developed with Japan's NICT, then goes through continued pre-training on Japanese finance corpora, SFT and DPO. Two points stand out. It writes its reasoning in Japanese before answering, which helps compliance staff check its basis. It also supports OpenAI-style function calling, and DPO used answers with too many or too few tool calls, wrong arguments, fabricated tool results or repeated calls as rejected samples. SFT ran on PFN's in-house MN-Core 2 AI chip. PFN says a coding-agent-centered porting method let engineers unfamiliar with the chip port the multi-board distributed training code in under a week, at training speed comparable to GPUs.

Benchmarks and Comparisons

All figures are vendor-reported, measured by PFN with vLLM, and not directly comparable with public leaderboards. On japanese-lm-fin-harness the average rose from 48.0 for the base to 55.8, ahead of Gemma 4 E4B (45.6) and gpt-oss-20b (43.0) but behind Qwen3.5-9B (60.1) and Ministral 3 14B (62.8). It scored 8.35 on the multi-turn pfmt-bench-fin-ja, close to Qwen3.5-9B's 8.53, and 71.3 on PFN's own Japanese finance tool-use benchmark, below Gemma 4 E4B (76.2). Independent third-party results have not been published.

Developer Hands-on Guide

Weights are at pfnet/plamo-3-8b-fin-instruct on Hugging Face; loading with transformers needs trust_remote_code, and the tokenizer needs numba. With vLLM 0.32.0 or later, add --reasoning-parser plamo3 and --tool-call-parser plamo3 to get an OpenAI-compatible API. Without self-hosting, the beta plamo-3.0-fin-beta is free to try on PLaMo API and Chat, with usage caps. PFN recommends passing only a few tools at a time (or filtering with a tool search step first), backing facts and numbers with retrieval tools and keeping a human reviewer; outputs are not investment advice. Companies and individuals under JPY 1 billion in annual revenue can use it commercially for free after registering.