🗾ITmedia AI+ (日本)•Freshcollected in 81m
Why PFN Is Building AI From Scratch
💡See why PFN rejects a shortcut approach and invests in building a domestic AI model from the ground up.
⚡ 30-Second TL;DR
What Changed
Preferred Networks is pursuing a domestically developed AI model under its PLaMo initiative.
Why It Matters
PFN’s approach highlights the trade-offs between using global foundation models and building a domestic model stack. For Japanese organizations, this strategy could be relevant where language quality, data control, and local technological independence are priorities.
What To Do Next
Review PLaMo’s official model documentation and compare its Japanese-language benchmarks with the foundation models your team currently uses.
Who should care:Researchers & Academics
Key Points
- •Preferred Networks is pursuing a domestically developed AI model under its PLaMo initiative.
- •The company is committed to developing the model from scratch rather than simply adapting an existing foundation model.
- •The discussion focuses on the control, customization, and strategic advantages of in-house model development.
🧠 Deep Insight
AI-generated analysis for this event.
🔑 Enhanced Key Takeaways
- •Preferred Networks (PFN) utilizes its proprietary MN-Core supercomputing architecture, specifically designed for deep learning, to accelerate the training of PLaMo models.
- •The PLaMo initiative emphasizes high performance in Japanese language processing, addressing specific cultural and linguistic nuances often overlooked by Western-centric foundation models.
- •PFN has integrated PLaMo into its broader industrial solutions, targeting sectors like manufacturing, robotics, and biotechnology where specialized domain knowledge is critical.
- •The company maintains a focus on 'sovereign AI' principles, aiming to reduce dependency on foreign cloud infrastructure and proprietary model APIs for Japanese enterprises.
- •PFN has actively released smaller, efficient versions of PLaMo under open-source licenses to foster a domestic ecosystem and encourage third-party development.
📊 Competitor Analysis▸ Show
| Feature | PLaMo (PFN) | GPT-4o (OpenAI) | Claude 3.5 (Anthropic) |
|---|---|---|---|
| Primary Focus | Japanese Industrial/Domain | General Purpose | General Purpose |
| Infrastructure | MN-Core (In-house) | Azure (Cloud) | AWS/GCP (Cloud) |
| Customization | High (Full-stack control) | Moderate (API/Fine-tuning) | Moderate (API/Fine-tuning) |
| Japanese Proficiency | Native/Specialized | High (Multilingual) | High (Multilingual) |
🛠️ Technical Deep Dive
- Architecture: PLaMo models are based on the Transformer architecture, optimized for high-throughput training on PFN's MN-Core hardware.
- Hardware Acceleration: Utilizes MN-Core, a custom-designed processor optimized for matrix operations essential to deep learning, offering superior power efficiency compared to general-purpose GPUs.
- Training Data: Employs a curated dataset emphasizing high-quality Japanese text, including technical documentation, academic papers, and domain-specific industrial data.
- Deployment: Designed for flexible deployment, supporting both cloud-based API access and on-premises installation for clients with strict data sovereignty requirements.
🔮 Future ImplicationsAI analysis grounded in cited sources
PFN will achieve significant cost-efficiency advantages in model training by 2027.
The continued scaling of their proprietary MN-Core hardware reduces reliance on expensive, high-demand commercial GPU clusters.
PLaMo will become the standard for Japanese industrial AI applications.
The combination of domain-specific training and on-premises deployment options addresses the primary security and accuracy concerns of Japanese manufacturing firms.
⏳ Timeline
2014-03
Preferred Networks is established in Tokyo, Japan.
2020-07
PFN begins operation of the MN-3 supercomputer, powered by their custom MN-Core processors.
2023-11
PFN announces the PLaMo initiative and releases the first PLaMo-13B model.
2024-03
PFN releases PLaMo-13B-instruct, an instruction-tuned version of their foundation model.
2025-05
PFN expands PLaMo capabilities to support multimodal inputs for industrial robotics applications.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: ITmedia AI+ (日本) ↗