PFN introduces its domestic AI model PLaMo 3.1 Prime

Summary

PFN has introduced PLaMo 3.1 Prime, a domestically produced Japanese LLM. The model stands out for its on-premises 'Deep Research' capability (autonomous, multi-step research generating cited reports) and a free API layer. PFN positions the product as a sovereign AI option for corporate research workflows.

Why it matters

For Japanese companies unable to send sensitive data to foreign cloud APIs, a model offering on-premises operation, built-in Deep Research, and internal document integration fills a genuine compliance gap. PFN’s series of 35 benchmark tests, where it outperformed Qwen, gpt-oss, and Gemma in safety scores and preferred saying 'I don't know,' signals a deliberate, trust-focused positioning. The free API layer also makes it easier for Japanese developers to test a non-US and non-China model in production environments.

Highlights

• On-premises Deep Research can autonomously conduct research lasting from minutes to tens of minutes by combining internal documents with web sources.
• An average ~15% improvement in hallucination benchmarks compared to PLaMo 3.0 Prime; top safety scores in a 9-domain comparison.
• Speculative decoding provides a 1.7x speedup and 90% fewer output tokens in non-reasoning mode.

Critical take

The same benchmark table PFN highlights for safety shows that PLaMo 3.1 Prime lags behind Qwen, gpt-oss, and Gemma in STEM, coding, and web-search-assisted tasks. This reveals that the 'flagship' title relies more on a narrow trust/safety niche than on broad capability. Furthermore, the 15% hallucination improvement was measured only against PFN’s own previous version. Consequently, the actual difference compared to frontier models could not be quantitatively determined.