PFN releases PLaMo 3.1 Prime with web search and Deep Research

Preferred Networks released PLaMo 3.1 Prime on October 27, 2026, bringing a new version of its Japanese-developed flagship model to PLaMo Chat and its API. The technical announcement describes changes to the model itself alongside new web-search and Deep Research features. This is an available service upgrade, rather than a promise of a future launch.

The release matters to organizations evaluating Japanese-language AI and deployments under their own control. Its strongest practical question is not whether another chatbot can produce a polished report, but whether a team can connect research tools to appropriate sources and audit the resulting claims.

Research tools and model changes arrive together

PFN says Deep Research can plan a task, search repeatedly and assemble a report with citations. The model was trained on multi-step search and reasoning tasks, with additional work on recognizing insufficient evidence and citing supporting material. A hybrid design also lets the same model operate with or without extended reasoning.

The company reports faster inference using speculative decoding and improvements on its evaluations. Those results are developer-reported, not tests performed by PromptsZone. A faster generation rate is not the same as completing every research assignment faster, and improved benchmark scores do not mean factual errors have disappeared.

Cloud access and on-premises deployment are different choices

According to the October 27 release notice, PLaMo 3.1 Prime is offered through Chat, API and on-premises arrangements. The Standard API price is listed as ¥60 per million input tokens and ¥250 per million output tokens. Enterprise on-premises and Provider plans require individual quotations; this announcement should not be read as an unrestricted public-weights release.

PFN also describes connecting on-premises Deep Research to internal search systems, including research limited to documents inside a closed environment. That is a deployment option, not evidence that every installation is isolated or ready to handle confidential information without configuration and review.

Check the free tier's data terms

The official product page says Free-plan inputs and outputs are used for training, while Standard and Provider plans are not. This distinction deserves attention before anyone pastes company documents into a trial. A free account is not automatically an appropriate evaluation environment for private material.

The launch includes a temporary Chat promotion allowing two free Deep Research uses per day through November 30, 2026. Usage limits still apply. Teams comparing the service should check current account eligibility, search-related charges and deployment terms directly rather than extrapolating from headline token prices.

A useful evaluation would start with public documents and questions whose answers are already known. Review whether each citation supports the surrounding claim, record unsupported conclusions and compare the time needed for human correction. Those checks are more informative than treating a fluent report as a finished research product.

Photo: Cabinet Public Affairs Office, Japan, via Wikimedia Commons, CC BY 4.0, unmodified. Archival AI discussion on May 9, 2023, including PFN's Daisuke Okanohara, second seated person from the right; not the PLaMo 3.1 launch or an endorsement.