Fixstars adds GLM-5.3 to its Japan-hosted inference service

Fixstars added GLM-5.3 to Fixstars Inference Cloud on November 20, making Z.ai’s existing open-weight model available through an API that runs on servers in Japan. It is available to customers with direct Fixstars contracts at launch. The company plans to add it to the FastMetal gateway later.
Fixstars uses a four-bit version of the model to reduce memory and compute requirements. The company lists prices per million tokens at ¥10.49 for input, ¥210 for output, and ¥10.34 for cache reads. It reported a median time to first token of 1.127 seconds in a Tokyo test with 64 concurrent requests and a cache hit rate of about 96%. That result does not establish the same latency for requests without cached input.
Fixstars says its inference servers and storage are in Japan and that it does not keep prompts or outputs in logs or use them for training.
The addition gives companies that need to process source code or documents within Japan another hosted model option. It is a new deployment of GLM-5.3, not a new release of the model itself.
Sources and citations
Revision history
Published .