Preferred Networks Launches PLaMo 3.1 Prime With Web Search and a Deep Research Mode That Can Run On-Premises

Preferred Networks (PFN) on October 27 released PLaMo 3.1 Prime, the latest version of its Japanese-developed large language model, adding web search and a Deep Research function that investigates a question across many sources and assembles a report, the company announced. PFN says Deep Research can run not only in the cloud but also on-premises and inside closed networks, a point aimed squarely at government bodies and companies that cannot send sensitive material to outside services.

According to PFN, PLaMo 3.1 Prime improves on PLaMo 3.0 Prime by about 15 per cent on average across eight hallucination benchmarks. A new non-reasoning mode cuts output tokens by roughly 90 per cent for tasks that do not need step-by-step thinking, and speculative decoding makes generation about 1.7 times faster. The base model comes from pretraining carried out in joint research with the National Institute of Information and Communications Technology (NICT), which also supported the safety evaluation.

Pricing is designed to widen use. PFN has added a free API tier, while the standard plan costs 60 yen per million input tokens and 250 yen per million output tokens. Users of the PLaMo Chat service can try Deep Research free twice a day until November 30.

What it means in practice

PLaMo is already part of public-sector work. An earlier version, PLaMo 2.0 Prime, was selected for the Digital Agency’s government AI environment, which recently opened AI agents to 180,000 government staff in Gennai. The model also sits inside QommonsAI, a service used by around 800 local governments, as well as Tachyon and miibo, and is offered through Amazon Bedrock Marketplace, Snowflake and Microsoft Foundry. PFN and NICT are also working with Sakura Internet on an active generative AI evaluation platform now being rolled out.

For a municipal office or a manufacturer with strict data rules, an on-premises research agent changes what is possible: staff can ask the model to compile background on a policy or a technical problem without documents leaving the building. The trade-off is that web results and multi-step reports bring new ways to be wrong, so the claimed hallucination gains will need to hold up in daily use, not only on benchmarks.

What to watch next is whether the Digital Agency and local governments move to the 3.1 generation, and how the free tier affects take-up among smaller Japanese developers who have so far defaulted to overseas models.