Japanese company Preferred Networks (PFN) views developing its PLaMo language model from scratch as a way to control the fundamental decisions involved in building the model, rather than merely attempting to compete with American and Chinese models on performance metrics. According to 田中大輔, head of the company's large language model development division, the importance of this approach lies in PFN's direct involvement in determining the design, data, and processing mechanisms from the outset.
This was stated in an interview with 田中大輔 and ムジビヤ・アディヤン, head of large language model planning and the person responsible for PreferredAI, during a discussion about the limits of describing artificial intelligence models as “local.” The term has no standardized definition: some companies rely on open models developed by external organizations and then add training focused on the Japanese language or specific fields, while PFN chose to build PLaMo using a “full scratch development” approach (Full Scratch).
Data Control Enhances Interpretability
PFN links in-house development to a clearer ability to assume responsibility toward customers. In many open models, only some information is available, such as the parameters that affect outputs, while training data and development methods remain undisclosed in full. When an open model is retrained on additional data, identifying the source of a particular behavior or problem in the outputs becomes more difficult.
In the case of PLaMo, the company says it can select the datasets used in training and understand their characteristics to a greater extent. This point is particularly important to companies dealing with intellectual property and needing to answer questions about the possibility that copyright-infringing data exists among the training materials. PFN also believes that knowledge of the training stages helps it investigate the causes of errors, even though some aspects of generative models remain impossible to fully explain.
田中 gives hallucinations as an example: if PLaMo provides an incorrect answer, the company may be able to link the problem to a specific part of the dataset and work to correct it. In a model built on top of an open model, however, identifying the cause may not always be possible.
Design Geared Toward the Japanese Language
According to PFN, the benefit of developing from scratch is not limited to transparency. The company designed PLaMo while taking into account knowledge associated with Japanese culture and characteristics, and developed its own linguistic analyzer, known as a Tokenizer, to convert text into units the model can process.
PFN says that when processing Japanese, this analyzer consumes approximately 20 to 30% fewer tokens than analyzers used by external models. The number of tokens affects processing efficiency and cost, so the company believes that improving this aspect can give local models a practical advantage, even when they face competition from models with strong general capabilities.
Owning the entire training cycle also makes it possible to prepare datasets linked to specific objectives. This may help when building specialized models for fields such as the financial sector, as it becomes easier to identify the type of data needed to enhance a particular performance area.
The Sovereignty Advantage Versus Development Costs
PFN places the development of PLaMo within a broader context involving geopolitical risks and the growing reliance on artificial intelligence as infrastructure. The company pointed out that a decision by the U.S. government led to the temporary suspension in June of access to the Fable 5 and Mythos 5 models from U.S. company Anthropic, considering that such developments illustrate the extent to which services are affected by international decisions.
However, full development is not the cheapest or fastest option. It requires more time and training resources than adding training to an open model, and may also make it more difficult to keep up with the latest open models in terms of performance. PFN therefore does not rule out using both approaches together; a retrained open model may be sufficient for some applications, and the company mentioned a use case in the medical field.
From PLaMo to Local Infrastructure
PFN provides its models through an application programming interface (API) as well as in local environments at the customer's premises (On-Premises), which suits organizations that prioritize managing their data. The company wants to strengthen PLaMo's presence by connecting it with external tools and applications on computers, and by developing its ability to perform long tasks autonomously and handle complex Japanese instructions.
The series includes the PLaMo 3.0 Prime model, in which the company focuses on combining high Japanese-language performance with low cost, in addition to PLaMo翻訳, which is dedicated to translation. 田中 says the goal is to reach a model or service that offers Japanese users the best cost efficiency, rather than merely increasing reasoning capability.
The computing infrastructure itself remains a challenge. PFN uses computing resources inside Japan as well as abroad, because domestic resources alone are insufficient to compete with current training requirements. In parallel, the company is developing dedicated artificial intelligence chips and pursuing vertical integration that could support the localization of computing resources in the future.
In July, PFN announced its cooperation with Japanese artificial intelligence company Noetra, backed by investments from major Japanese companies. PFN CEO 大輔岡野原 will lead model development, with 田中 and a number of engineers being dispatched to Noetra. The project initially aims to build a physical artificial intelligence model capable of operating robots autonomously, using computing infrastructure belonging to local operators. According to 田中's estimate, a model that could be described as “truly local” may appear as early as 2027.
What PFN's experience reveals is that a model's “local” status is not determined by performance alone. It is a trade-off between control over data and design, the ability to interpret errors, and language efficiency on the one hand, and training costs and the speed of access to the latest capabilities on the other. For this reason, the company's practical strategy appears closer to choosing the appropriate level of independence for each use case rather than treating development from scratch as the only solution.