Fixstars and Macnica unveiled a framework to support the deployment of Tenstorrent products in Japan. Their joint PoC demonstrated a 31.3% reduction in large language model inference time.
On 2 October 2026, Fixstars and Macnica announced that they had established a framework to provide domestic companies with end-to-end support for Tenstorrent products, from hands-on evaluation and performance optimization through deployment and operations. The two companies jointly conducted a proof of concept and demonstrated significant technical advantages.
The joint PoC used Tenstorrent's TT-QuietBox workstation to optimize the inference performance of the open-weight large language model gpt-oss-120b. The companies indicated that matrix multiplication for the MoE component, a bottleneck in decoding, was accelerated by 3.68x, reducing TPOT from 144.85 milliseconds to 99.49 milliseconds.
Under the integrated framework based on hands-on hardware validation, corporate customers can test their own models in their own environments before purchase and make quantitative performance assessments. Combining Macnica's AI infrastructure development capabilities with Fixstars' expertise in improving software implementations, the companies will provide one-stop support from identifying bottlenecks through operations.
Going forward, the companies plan to expand the scope to models other than gpt-oss-120b and to a variety of workloads, supporting companies in selecting and using AI infrastructure. The significance of the collaboration has been reported through Fixstars' public relations materials and media reports.