PFN releases PLaMo 3.1 Prime with on-premises Deep Research and a free API plan; Standard tier is 60 yen per 1M input to

Preferred Networks (PFN) began offering PLaMo 3.1 Prime on October 27, the flagship of its PLaMo model family, which it develops from scratch in Japan. The model ships with web search and a Deep Research function that can run in on-premises environments. PFN also opened a free plan for the PLaMo API. The details below come from a Japanese-language report by ITmedia (ITmedia AI＋) published October 28, 2026.

Deep Research, including on-premises

According to PFN, Deep Research takes a research topic, then plans the investigation and produces a report with citations autonomously, in minutes to tens of minutes. When deployed on-premises and connected to a company's internal document search system, it can combine internal documents with web information or restrict research to internal documents only.

Training and factuality

PLaMo 3.1 Prime is built on a pretrained model obtained through joint research with Japan's National Institute of Information and Communications Technology (NICT). PFN says post-training improved two behaviors: factuality (answering "I don't know" when it does not know) and faithfulness (reading supplied context correctly).

PFN reports that scores on eight hallucination-related benchmarks, including AA Omniscience, improved by about 15% on average over the previous PLaMo 3.0 Prime.

PFN also says that where 3.0 Prime refused questions about crime and other inappropriate actions, 3.1 Prime helps users make appropriate judgments, for example by explaining errors in the question's premise.

Benchmarks

PFN ran 35 benchmark tests against its own and competing models and published averages across nine domains. In both non-reasoning and reasoning modes, PFN's safety score was the highest, ahead of Alibaba's Qwen 3.8 27B, OpenAI's gpt-oss-120B and Google's Gemma 4 31B, among others. Performance on STEM, coding and web-search-assisted answers was below the competing models.

Speed and token usage

PFN says speculative decoding, in which a lightweight model predicts upcoming text and the main model verifies it, raised processing speed to 1.7 times that of 3.0 Prime. In non-reasoning mode, PFN says it cut output tokens by 90% while limiting quality loss, which it says reduces usage costs.

Availability and pricing

3.1 Prime is available from launch day through PLaMo Chat, the PLaMo API and on-premises deployments.

• API Free plan: free, with usage limits.
• API Standard plan: 60 yen per 1M input tokens, 250 yen per 1M output tokens.
• PLaMo Chat campaign: Deep Research is free up to twice per day through November 30.

What this means

The main change for enterprise users is that Deep Research is offered in an on-premises configuration that can search internal documents without the web. PFN's own benchmark summary shows a trade-off: it leads on safety but trails the compared models on STEM, coding and web-search-assisted answers. The hallucination and benchmark figures are PFN's own claims. The free API tier and the time-limited Deep Research campaign in PLaMo Chat give prospective users a way to test the model before committing to the Standard plan.