On June 22, 2026, the Tokyo-based AI startup Preferred Networks, Inc. (PFN) officially released PLaMo 3.0 Prime, a flagship large language model (LLM) developed entirely from scratch in Japan. Built on research conducted with Japan’s National Institute of Information and Communications Technology (NICT), this model aims to accelerate AI adoption in Japanese government agencies and critical infrastructure by offering high-level reasoning and superior cost-efficiency.

Achieving Advanced Reasoning and Practical Performance

Implementation of Japan’s First Reasoning Mode and Its Technical Background

PLaMo 3.0 Prime is a foundation model developed from the ground up by Preferred Networks (PFN), rather than being fine-tuned from a pre-existing overseas architecture. Incorporating insights gained from the beta version released in March 2026, PFN has formally implemented a “Reasoning Model” (also known as the “Long-thought” mode), making it the first Japanese domestic model capable of handling complex logical tasks. This Reasoning Model is specifically designed to organize multiple conditions and derive conclusions through a step-by-step process, making it ideal for mathematical problems, algorithmic challenges, and specialized decision-making support in corporate environments. The model configuration and characteristics of PLaMo 3.0 Prime are illustrated in the following diagram. . The development utilized pre-training models derived from joint research with the National Institute of Information and Communications Technology (NICT), which significantly enhanced the model’s ability to understand natural context and logical flow in the Japanese language. For applications requiring near-instantaneous responses, a “Non-reasoning model” is also provided, allowing businesses to choose between deep reasoning and rapid speed depending on their specific needs.

Context Expansion to 256K and AI Agent Capabilities

A major evolution for practical business use is the expansion of the context window, which represents the amount of information the model can process at once. The context length has been expanded fourfold from the beta version’s 64K tokens to 256K tokens (262,144 tokens) in the official release. This allows users to input several years’ worth of internal company documents, massive software source codes, or research reports spanning hundreds of pages for a single comprehensive analysis or summary. Furthermore, the model’s “tool calling” functionality has been strengthened to support its use as an autonomous “AI agent” that can interact with external systems and APIs. According to PFN’s internal evaluations, PLaMo 3.0 Prime demonstrates competitiveness with leading global closed-source models such as GPT-5.4 mini and Claude Haiku 4.5 in areas including Japanese instruction following, coding, and tool utilization. This advancement moves the technology beyond simple chatbots toward becoming a functional tool capable of handling integrated business processes.

Economic Efficiency Specialized for the Japanese Language and Infrastructure Strategy

Overwhelming Cost Advantage via a Unique Japanese Tokenizer

A decisive factor for Japanese corporate adoption is the overwhelming cost performance in processing Japanese text. Global LLMs are typically designed based on English, meaning a single Japanese character often consumes multiple tokens, which structurally increases operational costs for Japanese users. In contrast, PLaMo 3.0 Prime utilizes a unique tokenizer developed by PFN specifically for the Japanese language. This tokenizer processes Japanese text at an efficiency of approximately 0.54 tokens per character (roughly 1.85 characters per token). This is significantly more efficient than the GPT series (approx. 1.17 characters/token) or the Claude series (approx. 1.02 characters/token), effectively reducing the number of tokens consumed by 45% to 55% for the same amount of Japanese text. The following chart compares the processing costs per one million Japanese characters against other major models. . For the API “Standard” plan, the input fee is 60 yen and the output fee is 250 yen per one million tokens. When combined with the high token efficiency, the actual cost is reduced to a fraction of that of major overseas models.

Provisioning via Sakura Internet’s AI Inference Infrastructure

PFN is strengthening its collaboration with Japanese infrastructure providers to deliver its models. As of August 4, 2026, PLaMo 3.0 Prime became available on “Sakura’s AI Engine,” a generative AI inference API platform operated by Sakura Internet Inc. This service runs on domestic infrastructure equipped with NVIDIA H100 GPUs, offering a secure environment for Japanese companies concerned about data sovereignty and the risk of sensitive information being transferred overseas. Access to the model is provided on an application basis, and approved users are granted a dedicated environment. By combining PFN’s advanced software with robust domestic physical infrastructure, the model facilitates easier adoption in critical sectors such as finance, healthcare, and public services where extremely high security is required.

Proven Social Implementation and Governance Challenges in Japan

Expansion to 800 Municipalities and Promotion of Administrative DX

PLaMo 3.0 Prime is rapidly penetrating the administrative sector in Japan due to its high Japanese language performance and security. It is a standard component of “QommonsAI,” a generative AI platform for government agencies provided by Polimill Inc. As of June 2026, this solution has been adopted by approximately 800 local municipalities in Japan, contributing to administrative efficiency through tasks such as drafting responses for local assembly sessions. Furthermore, PLaMo was selected as one of the domestic models for trial in “GENNAI,” a generative AI environment developed by the Japanese Digital Agency. In the educational sector, Shiga University entered into an agreement to deploy “PLaMo Chat” and “PLaMo Translation” for all students and faculty members starting in April 2026. The following diagram illustrates the major adoption areas of domestic LLMs and PLaMo’s specific position in the market. . In the financial sector, PFN is also expanding into “PLaMo-Fin-Prime,” a model specialized with financial industry knowledge, designed to support high-level professional work while maintaining strict confidentiality.

Safety Evaluation and Risk Management for Corporate Adoption

To ensure secure corporate adoption, PLaMo 3.0 Prime has achieved high scores in safety evaluations conducted by third-party organizations. In the “HELM Safety” benchmark suite developed by the Stanford University Center for Research on Foundation Models (CRFM), PLaMo demonstrated safety performance equivalent to or higher than that of major global models, with enhanced considerations for sensitive information and avoiding the output of dangerous content. However, from the perspective of information systems departments, a cautious evaluation is still required. PFN reported a net loss of approximately 7.8 billion yen for the fiscal year ending June 2025 and underwent a significant capital reduction, moving from approximately 7.9 billion yen to 50 million yen. These financial fluctuations reflect the massive upfront investments typical of AI infrastructure development. Additionally, as of June 2026, PLaMo’s API service had not yet obtained SOC2 or ISMAP certifications. Companies considering adoption are advised to perform individual verifications against their own security requirements and carefully assess the vendor’s long-term business continuity.

The Current State of Japanese AI and Future Prospects

Core of a Hybrid Coexistence Strategy

Since 2026, the generative AI market in Japan has shifted from a pure performance competition against global giants toward a “Hybrid Coexistence Strategy” tailored to specific use cases. PLaMo 3.0 Prime is expected to play a central role in this strategy, where global models are used for general tasks while domestic models are utilized for Japanese-specific context, data sovereignty, and compliance with local regulations. In particular, its unique tokenizer efficiency and the logical adherence of the Reasoning Model create a competitive advantage rarely seen in other countries. Looking ahead, further cost reductions and improvements in power efficiency are expected if the model is integrated with PFN’s proprietary “MN-Core L1000” AI chips, potentially creating a vertically integrated Japanese AI foundation. As competition intensifies with other domestic players like Rakuten and SoftBank, the primary focus will be on how PFN can expand its strong track record in government and finance into the core systems of private enterprises.

References

• [Preferred Networks Releases Flagship Japanese Foundation Model PLaMo 3.0 Prime] https://www.preferred.jp/ja/news/pr20260622

• [PLaMo Platform Console] https://console.platform.preferredai.jp/

• [Digital Agency: Guidelines for Procurement and Utilization of Generative AI for Administrative Evolution and Innovation] https://www.digital.go.jp/assets/contents/node/basic_page/field_ref_resources/8b8d1635-cea3-4d36-bb9a-bac732779d1e/5107818b/20260612_meeting_executive_guideline_02.pdf

【#PFN #PLaMo #DomesticAI #GenerativeAI #LLM #ScienceAndTechnology #DX #Startup】