Preferred Networks (PFN), a Japanese AI company, released its latest generative AI foundation model, PLaMo 3.0 Prime, on June 22, 2026. This flagship model builds upon the PLaMo 3.0 Prime β version, incorporating feedback from corporate users to improve reasoning capabilities and overall stability.

The model features significantly increased reasoning performance, achieved by doubling the steps used in reinforcement learning compared to its predecessor. To better support AI agent applications, PFN extended the input context length from 64K to 256K. Additionally, the company introduced a specialized non-reasoning model optimized for use cases requiring high-speed responses.

Regarding safety, the model was trained using data provided by the National Institute of Information and Communications Technology (NICT). Tests using the Stanford University HELM Safety benchmark showed performance equal to or exceeding competitor models across six categories, including violence, fraud, and discrimination. The integration of a Structured Output feature further allows the LLM to force responses into specific data formats, simplifying integration with external APIs and systems.

In terms of benchmarks, PLaMo 3.0 Prime demonstrates strong performance against open models like Qwen3.6-27b and gpt-oss-120b, as well as closed models in the same price tier, such as GPT-5.4 Mini and Claude Haiku 4.5. While the model excels in instruction following, dialogue, tool use, medicine, and code generation, the company identified web search and mathematical reasoning as areas for future improvement. The model is currently available via PLaMo Chat and API, with a free plan included, as PFN continues its integrated strategy of developing everything from infrastructure to end-user solutions.

Preferred Networks (PFN), a Japanese AI company, released its latest generative AI foundation model, PLaMo 3.0 Prime, on June 22, 2026. This flagship model builds upon the PLaMo 3.0 Prime β version, incorporating feedback from corporate users to improve reasoning capabilities and overall stability.

The model features significantly increased reasoning performance, achieved by doubling the steps used in reinforcement learning compared to its predecessor. To better support AI agent applications, PFN extended the input context length from 64K to 256K. Additionally, the company introduced a specialized non-reasoning model optimized for use cases requiring high-speed responses.

Regarding safety, the model was trained using data provided by the National Institute of Information and Communications Technology (NICT). Tests using the Stanford University HELM Safety benchmark showed performance equal to or exceeding competitor models across six categories, including violence, fraud, and discrimination. The integration of a Structured Output feature further allows the LLM to force responses into specific data formats, simplifying integration with external APIs and systems.

In terms of benchmarks, PLaMo 3.0 Prime demonstrates strong performance against open models like Qwen3.6-27b and gpt-oss-120b, as well as closed models in the same price tier, such as GPT-5.4 Mini and Claude Haiku 4.5. While the model excels in instruction following, dialogue, tool use, medicine, and code generation, the company identified web search and mathematical reasoning as areas for future improvement. The model is currently available via PLaMo Chat and API, with a free plan included, as PFN continues its integrated strategy of developing everything from infrastructure to end-user solutions.