Eigen RadarAI
Archive
Analysis

PLaMo 3.1 Prime brings sourced research to closed company networks

Preferred Networks has released PLaMo 3.1 Prime with web search and a Deep Research function that plans investigations and produces reports with sources. Companies can run the research workflow on their own infrastructure, combining internal documents with web information or restricting it to documents inside a closed network. The release also changes how the model handles unknown answers and inappropriate requests.

Artificial Intelligence··Night
An upward view of a grey-brown stone office building corner, with deeply recessed windows extending along two facades and pale sky reflected in the glass.

Research moves from a question to a sourced report

Preferred Networks, a Japanese AI developer, began offering its PLaMo 3.1 Prime language model on October 27. The release adds web search and Deep Research, a function that prepares a research plan, gathers information and organizes the results into a report with sources. It is available through PLaMo Chat and the company’s application programming interface, which lets other software use the model.[1], [2]

Users supply a research topic, after which the system repeatedly collects, sorts and analyzes information. Preferred Networks identifies market research, competitor research and following technical developments as intended uses. Web search also allows ordinary answers to draw on current information and identify where it came from.[1]

Internal documents can stay inside a closed network

Deep Research can run on a company’s own infrastructure. Connecting an internal document-search system allows a research task to draw on both company documents and web information. In a closed network, the research can instead be restricted to internal documents. The deployment choice therefore also determines which information the research function can consult.[1]

Training changes address unknown answers and inappropriate requests

The model builds on pretraining work with Japan’s National Institute of Information and Communications Technology, a public research organization that also supported safety evaluation. Preferred Networks says post-training focused on acknowledging unknown answers and following the supplied context. For inappropriate requests, it illustrates responses that describe a mistaken premise. Users can select reasoning or non-reasoning modes; a smaller model also proposes text continuations for the main model to verify.[1]

References

  1. News sourcePreferred NetworksPLaMo 3.1 Prime adds research workflows for closed networks↩1↩2↩3↩4
  2. News sourceImpress WatchPFN releases PLaMo 3.1 Prime with Deep Research↩
AI safety and regulationThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →