樹林AI released JJ-1, an AI model for Japanese companies, as open source. It achieved an accuracy rate of 97.43% on a 6,033-question benchmark, delivers inference in 8–30 milliseconds and cuts inference costs by up to 99% compared with the previous approach.
樹林AI released its JJ-1 family of AI models specialized in operational decision-making at Japanese companies and began offering it free of charge as open source. JJ-1 reportedly achieved an accuracy rate of 97.43% on a 6,033-question benchmark covering Japanese business workflows, outperforming overseas comparison models.
Decision processing is reported to take 8–30 milliseconds, while inference costs can reportedly be reduced by up to 99% compared with conventional generative LLM methods.
As a sovereign AI developed in Japan, 樹林AI says it aims to develop AI suited to the Japanese language, Japanese business practices and corporate operations, and to return the technology to society. In line with the government's 2026 policy of open AI sovereignty, it advocates promoting self-reliance while making appropriate use of overseas technology rather than excluding it.
JJ-1 was evaluated using 6 datasets combining publicly available Japanese- and English-language data with proprietary data designed for Japanese business workflows. Its overall accuracy rate was 79.20%. On proprietary data covering Japanese business workflows, it recorded an accuracy rate of 97.43%, while other models reportedly remained in the low 91% range.
Rather than generating text one word at a time, JJ-1 directly evaluates defined options to determine the next action. This enables applications supporting corporate decision-making, including routing inquiries, assessing urgency, determining whether escalation is required, classifying information and selecting business workflows.