Skymizer Introduces HTX301 Inference Chip for Large Language Models

Ahead of COMPUTEX 2026, Skymizer announced in a press release the HTX301 inference chip, a hardware platform designed to enable ultra large language model inference on a single PCIe card. The chip is built on the company’s HyperThought architecture, which integrates hardware and software to optimize AI inference performance.

The HTX301 reference chip allows enterprises to run models with up to 700 billion parameters locally, using a single PCIe card powered by six chips and 384 GB of memory. The card operates at roughly 240 watts, eliminating the need for GPU clusters, high speed interconnects, or extensive cooling systems.

HyperThought can scale from a single chip to six chips per card, serving models from four to 700 billion parameters. The platform supports deployment across various environments, from edge devices to small data centers. It is based on Skymizer’s LISA instruction set architecture, optimized for transformer inference.

The company stated that the HTX301 simplifies infrastructure for on-premise AI workloads, improving power efficiency and data privacy. It manages inference phases through a unified software stack that separates compute intensive prefill operations from memory bandwidth intensive decode operations. Additional details about the HyperThought roadmap will be presented at COMPUTEX 2026.