Japanese AI Firm Preferred Networks Releases Open-Weight Financial LLM PLaMo 3.0 Fin for On-Premise Use
Preferred Networks released PLaMo 3.0 Fin, an 8-billion-parameter open-weight LLM specialized for finance, enabling banks to run AI locally to protect data privacy while offering competitive Japanese-language performance.
Japanese artificial intelligence company Preferred Networks (PFN) released "PLaMo 3.0 Fin," an open-weight large language model specifically designed for the financial industry, on November 9, 2026. The model features 8 billion parameters and aims to address the regulatory constraints faced by banks and securities firms that cannot transfer customer data to external clouds. By providing a lightweight model capable of running on local servers or in closed network environments, PFN offers an alternative to cloud-based generalist LLMs for financial institutions bound by strict data compliance rules.
PLaMo 3.0 Fin was developed through continued pre-training, supervised fine-tuning (SFT), and direct preference optimization (DPO) based on a foundation model co-developed by PFN and Japan's National Institute of Information and Communications Technology (NICT). Notably, the SFT phase utilized PFN's proprietary AI semiconductor, "MN-Core 2," as the computing infrastructure. PFN reported that code porting to MN-Core 2 was completed within one week using an AI coding assistant, achieving training speeds comparable to GPU environments. The model supports a context length of 262,144 tokens, includes reasoning capabilities with Japanese output for thought processes, and can call external tools such as web search.
In terms of performance, PFN's published benchmark results show that PLaMo 3.0 Fin achieved an average score of 55.8 across five tasks in the "japanese-lm-fin-harness" test set, an improvement of 7.8 points over the base model. This score surpasses Gemma 4 E4B (45.6) and gpt-oss-20b (43.0). However, it remains behind similarly sized models like Qwen3.5-9B (60.1) and Ministral 3 14B (62.8). In text generation quality assessed by "pfmt-bench-fin-ja," the model scored 8.35, and its tool-calling capability averaged 71.3 points, both lower than the current leader, Gemma 4 E4B. These results are self-reported by the vendor and have not yet been independently reproduced.