Preferred Networks (PFN) has released the beta version of its new flagship model, PLaMo 3.0 Prime. This release marks a significant milestone as the first model developed from scratch in Japan to possess integrated reasoning capabilities. Unlike traditional models that generate answers directly, PLaMo 3.0 Prime can output its 'Thinking Process,' allowing it to organize its thoughts step-by-step and verify logic before delivering a final answer.

To achieve this breakthrough, PFN completely overhauled its training pipeline. In addition to conventional Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), the company introduced Reinforcement Learning (RL) into the model's development. This approach resulted in benchmark scores, such as on the AIME 2024 mathematical reasoning test, that significantly outperform the previous PLaMo 2.2. Furthermore, by utilizing a technique known as YaRN during continued pre-training, the team successfully doubled the model's context length to 64K tokens.

In practical demonstrations, the model has shown it can solve intricate logic problems, such as the classic 'river crossing puzzle,' by evaluating conditions and presenting a flawless sequence of steps. While much of the thinking process is currently generated in English, PFN plans to enable reasoning in Japanese to further enhance processing efficiency and accuracy. By leveraging its 'vertically integrated' strengths—spanning from proprietary chips and infrastructure to model development—PFN aims to further expand context lengths and reasoning depth. The company is currently seeking corporate monitors to accelerate the industrial application of generative AI in Japan.

Preferred Networks (PFN) has released the beta version of its new flagship model, PLaMo 3.0 Prime. This release marks a significant milestone as the first model developed from scratch in Japan to possess integrated reasoning capabilities. Unlike traditional models that generate answers directly, PLaMo 3.0 Prime can output its 'Thinking Process,' allowing it to organize its thoughts step-by-step and verify logic before delivering a final answer.

To achieve this breakthrough, PFN completely overhauled its training pipeline. In addition to conventional Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), the company introduced Reinforcement Learning (RL) into the model's development. This approach resulted in benchmark scores, such as on the AIME 2024 mathematical reasoning test, that significantly outperform the previous PLaMo 2.2. Furthermore, by utilizing a technique known as YaRN during continued pre-training, the team successfully doubled the model's context length to 64K tokens.

In practical demonstrations, the model has shown it can solve intricate logic problems, such as the classic 'river crossing puzzle,' by evaluating conditions and presenting a flawless sequence of steps. While much of the thinking process is currently generated in English, PFN plans to enable reasoning in Japanese to further enhance processing efficiency and accuracy. By leveraging its 'vertically integrated' strengths—spanning from proprietary chips and infrastructure to model development—PFN aims to further expand context lengths and reasoning depth. The company is currently seeking corporate monitors to accelerate the industrial application of generative AI in Japan.