Key Findings
PFN has launched PLaMo 3.1 Prime, a domestic Japanese generative AI model designed to automate corporate research tasks. The primary addition is the Deep Research function, which autonomously handles the entire research workflow: planning, information collection, organization, analysis, and the creation of reports with citations. This process, which previously took human operators hours or days, is now completed in several minutes to tens of minutes.
Technical Details
The model introduces several architectural and functional optimizations:
- Factuality and Reliability: Post-training has improved the model’s ability to admit when it does not know an answer rather than guessing. Hallucination-related benchmark scores improved by an average of approximately 15% compared to PLaMo 3.0 Prime.
- Inference Modes: A “non-inference mode” reduces output tokens by 90% while maintaining quality, lowering costs and latency. A “reasoning mode” remains available for complex tasks.
- Performance: By implementing Speculative Decoding—where a lightweight model predicts text and the main model verifies it—inference speed is 1.7 times faster than PLaMo 3.0 Prime.
- Safety: Revised training pipelines and data, supported by the National Institute of Information and Communications Technology (NICT), allow the model to explain why certain actions are incorrect rather than simply refusing to answer.
Background and Context
The development addresses two primary barriers to AI adoption in corporate environments: the inability of models to access real-time information (leading to hallucinations) and security concerns regarding the use of cloud-based AI from overseas providers for sensitive data. PLaMo 3.1 Prime mitigates these by integrating web search for latest information and offering on-premises deployment, which allows for research limited to internal company documents.
Significance and Outlook
PLaMo 3.1 Prime is available via PLaMo Chat, PLaMo API, and on-premises installations. The API offers a free plan and a Standard plan priced at 60 yen per million input tokens and 250 yen per million output tokens.
Source: https://www.watch.impress.co.jp/docs/news/2151518.html
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
