Preferred Networks (PFN) has released PLaMo 2.2 Prime, the latest iteration of its domestic AI foundation model.
By substantially increasing post-training data based on internal and external feedback, PFN has refined the model’s ability to follow complex user prompts—a capability known as instruction following—and deepened its expertise in the high-stakes medical domain.

A key technical milestone was the improvement in IFBench, a benchmark measuring an LLM's ability to adhere to constraints like word limits and specific formatting.
The new model achieved a score of 37.8%, up nearly 10% from its predecessor's 29.0%.
Notably, PLaMo achieves this high compliance with fewer generated tokens compared to traditional reasoning models that output their entire thought processes in text.
To better evaluate Japanese language nuances, PFN also developed and released JFBench, where the model demonstrated performance on par with global frontier models.

The update also focused on multi-turn role-play, where the AI maintains a consistent persona and context over multiple exchanges.
Utilizing datasets from PFN's AI interview service, Talent Scouter, the model saw a performance jump of over 15% in this area.
This allows chatbots to handle practical customer interactions more naturally while strictly adhering to specific rules, such as limiting every response to exactly three sentences.

In the medical sector, the model was tested against MedRECT-ja for detecting clinical documentation errors and the Japan Medical Licensing Examination (JMLE).
It reached an impressive accuracy rate of 70.7% on past medical exam questions.
While there is still a gap compared to the world's top-tier frontier models, PLaMo shows steady evolution.
PFN plans to lead Japanese AI development through its vertical integration strategy, leveraging its proprietary MN-Core hardware and specialized software stack to d

Preferred Networks (PFN) has released PLaMo 2.2 Prime, the latest iteration of its domestic AI foundation model.
By substantially increasing post-training data based on internal and external feedback, PFN has refined the model’s ability to follow complex user prompts—a capability known as instruction following—and deepened its expertise in the high-stakes medical domain.

A key technical milestone was the improvement in IFBench, a benchmark measuring an LLM's ability to adhere to constraints like word limits and specific formatting.
The new model achieved a score of 37.8%, up nearly 10% from its predecessor's 29.0%.
Notably, PLaMo achieves this high compliance with fewer generated tokens compared to traditional reasoning models that output their entire thought processes in text.
To better evaluate Japanese language nuances, PFN also developed and released JFBench, where the model demonstrated performance on par with global frontier models.

The update also focused on multi-turn role-play, where the AI maintains a consistent persona and context over multiple exchanges.
Utilizing datasets from PFN's AI interview service, Talent Scouter, the model saw a performance jump of over 15% in this area.
This allows chatbots to handle practical customer interactions more naturally while strictly adhering to specific rules, such as limiting every response to exactly three sentences.

In the medical sector, the model was tested against MedRECT-ja for detecting clinical documentation errors and the Japan Medical Licensing Examination (JMLE).
It reached an impressive accuracy rate of 70.7% on past medical exam questions.
While there is still a gap compared to the world's top-tier frontier models, PLaMo shows steady evolution.
PFN plans to lead Japanese AI development through its vertical integration strategy, leveraging its proprietary MN-Core hardware and specialized software stack to d