Fixstars starts running the GLM-5.3 language model in Japan
Japan’s Fixstars is now offering the GLM-5.3 language model through a cloud service that handles requests and storage inside Japan. That gives companies facing restrictions on sending code and documents abroad a domestic option for using the model. Fixstars adapted weights originally released by China’s Z.ai and says it discards inputs and outputs after each request.
Artificial Intelligence··Evening
The cloud service keeps processing in Japan
Japanese software company Fixstars has added the GLM-5.3 language model to Fixstars Inference Cloud, its service for running artificial intelligence models remotely. The computers that process requests and the storage systems used by the service are in Japan. Customers can send requests through an interface that works with applications designed for OpenAI’s service.[1], [2]
Fixstars compressed the published weights
China’s Z.ai released the model’s weights in August. Those numerical values determine how a trained model processes information. Fixstars converted them to a four-bit format created by chip designer Nvidia and adapted the model for its service. Fixstars says the smaller representation reduces the memory and computing work needed to run the model, which Z.ai designed for tasks including software development.[1]
Direct customers get access first
The first customers able to use the model are those with a direct contract. Fixstars says it operates and maintains the infrastructure itself, including the work needed to adapt models for its own servers. The service is aimed in part at companies that face limits on sending source code and internal documents overseas. Fixstars says it discards questions and answers after processing and keeps them out of its training data. It also plans access through FastMetal, a gateway for connecting to models, at a later date.[1]