Preferred Networks Launches PLaMo-2 Open LLM

Preferred Networks (PFN), one of Japan's most prominent AI startups, has officially launched PLaMo-2, its next-generation open-weight large language model designed to deliver state-of-the-art performance in Japanese language processing. The release marks a significant step forward in the global push to build high-quality, non-English LLMs that can compete with Western-developed models like Meta's Llama 3 and Mistral's offerings.

PLaMo-2 arrives at a time when the AI industry is increasingly recognizing that English-centric models fall short for billions of users worldwide. PFN's latest model aims to bridge that gap for Japanese speakers while also maintaining strong multilingual capabilities.

Key Takeaways at a Glance

• PLaMo-2 is an open-weight large language model built by Preferred Networks, headquartered in Tokyo
• The model is optimized for Japanese language understanding and generation, outperforming many existing Japanese LLMs on key benchmarks
• PFN trained PLaMo-2 on a massive multilingual corpus with a heavy emphasis on high-quality Japanese text data
• The model is released under an open license, allowing researchers and developers to fine-tune and deploy it
• PLaMo-2 builds on PFN's earlier PLaMo-1 model, incorporating architectural improvements and expanded training data
• The launch positions PFN as a leading competitor in the Japanese open-source AI ecosystem alongside companies like NEC, Sakana AI, and CyberAgent

PFN Targets the Japanese AI Gap With PLaMo-2

Preferred Networks has long been recognized as one of Japan's foremost deep learning companies, with roots in robotics, drug discovery, and industrial optimization. The company's pivot toward large language models reflects a broader strategic shift across the Japanese tech sector, where firms are racing to build sovereign AI capabilities.

PLaMo-2 represents a substantial upgrade over its predecessor. The original PLaMo-1, released in late 2023, demonstrated that a Japanese company could build competitive foundation models. However, it faced limitations in reasoning tasks and struggled with longer-context scenarios compared to leading Western models.

With PLaMo-2, PFN has addressed many of those shortcomings. The model incorporates architectural enhancements that improve both efficiency and output quality. Training was conducted on PFN's proprietary compute infrastructure, which includes clusters of NVIDIA GPUs optimized for large-scale model training.

Technical Architecture and Training Details

While PFN has not disclosed every architectural detail, several key technical aspects of PLaMo-2 set it apart from competing Japanese LLMs.

The model reportedly uses a transformer-based architecture with modifications designed to handle the unique challenges of Japanese text. Japanese writing combines 3 scripts — hiragana, katakana, and kanji — and requires specialized tokenization strategies that differ significantly from English-focused approaches.

PLaMo-2's training data pipeline is another differentiator. Key technical highlights include:

• A custom tokenizer optimized for Japanese morphological analysis, reducing token counts and improving inference efficiency
• Training on a curated corpus that balances Japanese web data, academic papers, government documents, and translated high-quality English content
• Support for extended context windows, enabling the model to process longer documents common in Japanese business and legal settings
• Alignment tuning using both RLHF (Reinforcement Learning from Human Feedback) and DPO (Direct Preference Optimization) techniques
• Competitive performance on Japanese benchmarks such as JGLUE, JCommonsenseQA, and MARC-ja

Compared to Meta's Llama 3, which excels primarily in English and a handful of European languages, PLaMo-2 delivers meaningfully better results on Japanese-specific tasks. This advantage stems from the sheer volume of Japanese training data and the linguistically informed tokenization strategy PFN employs.

How PLaMo-2 Stacks Up Against Competitors

The Japanese LLM landscape has grown increasingly crowded over the past 18 months. Several major players have entered the market, each with different approaches and target audiences.

NEC's cotomi model targets enterprise customers with a focus on security and on-premises deployment. CyberAgent's CALM series has gained traction in advertising and content generation. Meanwhile, Sakana AI, co-founded by former Google Brain researchers, has taken a more research-oriented approach with novel evolutionary model merging techniques.

PLaMo-2 differentiates itself through its combination of open weights and strong benchmark performance. Many enterprise-focused Japanese LLMs remain closed-source, limiting their utility for academic researchers and smaller companies. PFN's decision to release PLaMo-2 under an open license could accelerate adoption across Japan's startup ecosystem and university research labs.

International models also compete for Japanese users. OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet both handle Japanese reasonably well, but they operate as closed APIs with data sovereignty concerns that make some Japanese enterprises hesitant to adopt them for sensitive workloads.

Why Japanese-Optimized LLMs Matter Globally

Japan represents the world's 4th-largest economy and a massive potential market for AI services. Yet Japanese language processing has historically been underserved by the global AI community. The structural complexity of written Japanese — with its mixed scripts, context-dependent honorifics, and ambiguous word boundaries — makes it one of the more challenging languages for LLMs to master.

The launch of PLaMo-2 matters beyond Japan for several reasons. First, it contributes to the growing movement toward linguistic diversity in AI. As models like Llama, Mistral, and Qwen demonstrate strong performance in their respective language families, PLaMo-2 adds another high-quality option for a language spoken by over 125 million people.

Second, PFN's approach to training and releasing an open model could serve as a blueprint for other nations seeking to develop sovereign AI capabilities. Countries across Asia, the Middle East, and Europe are investing heavily in domestic LLMs, and PLaMo-2 offers a case study in how a relatively small company can punch above its weight.

Third, the model's open-weight nature means it can be fine-tuned for specialized applications. Japanese industries like automotive manufacturing, healthcare, legal services, and financial compliance all have unique terminology and documentation requirements that benefit from domain-specific adaptation.

What This Means for Developers and Businesses

For developers working with Japanese text, PLaMo-2 opens up new possibilities. The open-weight release means teams can download the model, run it on their own infrastructure, and customize it for specific use cases without relying on external API providers.

Practical applications likely to benefit include:

• Customer service automation for Japanese enterprises that need natural, polite conversational AI
• Document summarization of lengthy Japanese legal and regulatory texts
• Translation workflows that require deep understanding of Japanese context and nuance
• Code generation with Japanese comments and documentation
• Content creation for Japanese media, marketing, and e-commerce platforms

Businesses concerned about data residency — a significant issue in Japan, where regulations around personal information protection have tightened — can deploy PLaMo-2 entirely on-premises, eliminating the need to send sensitive data to overseas cloud providers.

The model's availability also lowers the barrier to entry for Japanese AI startups. Instead of spending millions of dollars training a foundation model from scratch, small teams can fine-tune PLaMo-2 for their specific vertical and bring products to market faster.

Looking Ahead: PFN's Roadmap and the Future of Japanese AI

Preferred Networks has signaled that PLaMo-2 is not the end of its LLM ambitions. The company is expected to continue iterating on the model, with potential future releases incorporating multimodal capabilities — the ability to process images, audio, and video alongside text.

The Japanese government has also been a tailwind for domestic AI development. Japan's Ministry of Economy, Trade and Industry (METI) has allocated significant funding to support AI infrastructure buildouts, including subsidies for GPU clusters and data center construction. PFN stands to benefit from these investments as it scales its training capabilities.

Looking at the broader landscape, the release of PLaMo-2 reinforces a clear global trend: the era of English-only AI dominance is ending. From Alibaba's Qwen models optimized for Chinese to France's Mistral championing European AI sovereignty, regional LLMs are becoming a critical layer of the global AI stack.

PFN's PLaMo-2 ensures that Japan has a credible, open, and technically competitive entry in this race. Whether it can sustain that momentum against both domestic rivals and well-funded international competitors will depend on continued investment, community adoption, and the speed at which it delivers its next generation of improvements.

For now, developers and researchers interested in Japanese language AI have a powerful new tool at their disposal — and the open-weight approach means the community can help shape where PLaMo-2 goes next.