2 TB card-memory configuration is designed to fit Kimi K3 within one server for enterprise on-premises deployment

TAIPEI, Nov. 5, 2026 /PRNewswire/ -- Skymizer and MiTAC Computing today announced a joint showcase at SC26 featuring Moonshot AI's Kimi K3—a 2.8-trillion-parameter mixture-of-experts model—running on a single MiTAC server equipped with eight Skymizer HTX301 LPU cards and 2 TB of card-attached memory.

Kimi K3 uses native MXFP4 (quantization-aware trained) weights. The 2 TB memory capacity is designed to accommodate the approximately 1.56 TB weight set while leaving capacity for KV cache. Final memory allocation and supported workloads remain subject to engineering validation.

Skymizer will also participate in ICCAD-Expo 2026 in Beijing on November 19–20 with local partners, bringing the same HTX301 product announcement to audiences in China. Together, the SC26 announcement in Chicago and participation in Beijing give the product announcement a global reach, connecting with enterprise AI customers and partners in both markets.

From COMPUTEX to a larger model platform

The platform builds on Skymizer's demonstration of Qwen3-VL-30B-A3B, a 30-billion-parameter multimodal MoE model, at COMPUTEX 2026 on June 2–5.

"At COMPUTEX, we demonstrated Qwen3-VL-30B-A3B on our evaluation hardware," said Luba Tang, Chief Technology Officer of Skymizer. "Our team is currently successfully extending that work to a single-server configuration designed for Kimi K3. LISA™ provides a common instruction set and toolchain as we develop the memory-residency and scheduling capabilities required for larger models."

"Enterprises want greater control over the models, data, and infrastructure behind their AI services," said William Wei, Chief Marketing Officer of Skymizer. "Our work with MiTAC aims to make larger open-weight models more practical to deploy on premises. Workload validation remains essential to assessing power, cost, and deployment requirements."

"At SC26, MiTAC Computing is showcasing scalable, full-stack AI infrastructure solutions designed to support evolving AI workloads. Our collaboration with Skymizer explores Prefill–Decode disaggregation and optimized AI inference deployment across diverse workloads. By combining MiTAC's server engineering expertise and scalable infrastructure platforms with Skymizer's inference technologies, we aim to provide customers with greater flexibility and efficiency in AI deployment," said a MiTAC spokesperson.

Built for enterprise on-premises deployment

For enterprises seeking to operate AI within their own infrastructure, fitting a large open-weight model within one air-cooled server offers a practical direction for deployment. Skymizer's capacity-first cards and decode-optimized architecture are designed to support that goal while allowing enterprises to retain control over their models, private data, and AI operations. Deployment still requires careful matching of memory, compute, power, and cooling to the workload. The approach targets enterprise needs in financial services, healthcare, government, manufacturing, and semiconductor design, where data governance and operational control are central to adoption.

AI-assisted delivery for enterprise deployment

Bringing ultra-large models on premises also requires a practical path from hardware to a working service. Skymizer is developing an AI-assisted delivery approach to help enterprises select models for their workloads, prepare the software environment, install and configure the inference stack, and validate the deployment before use. The aim is to make these steps more consistent and repeatable while keeping model, data, and deployment decisions under enterprise control.

The approach is designed around shared deployment tools that different AI assistants can use, rather than dependence on a single agent or service. Over time, Skymizer plans to extend this assistance to ongoing operations and maintenance, helping teams diagnose issues and manage updates while retaining human oversight and approval. This work supports Skymizer's "Own Your Intelligence" direction by addressing how enterprises deploy and maintain the AI systems they choose to own.

Availability

HTX301 chip production is expected to begin in Q1 2027. Development and validation of the server platform are ongoing. Sampling and broader commercial availability will be announced separately as integration and validation progress. Further information: skymizer.ai/events/sc26 

About MiTAC Computing Technology Corp.

MiTAC Computing Technology Corp., a subsidiary of MiTAC Holdings, delivers comprehensive, energy-efficient server solutions backed by industry expertise dating back to the 1990s. Specializing in AI, HPC, cloud, and edge computing, MiTAC Computing employs rigorous methodologies to ensure uncompromising quality — across barebones, systems, racks, and cluster levels — fully achieving performance and integration.

About Skymizer

Founded in 2013, Skymizer Taiwan Inc. develops purpose-built Language Processing Unit (LPU) IP and silicon for on-premises and on-device LLM inference. Its HyperThought™ platform pairs compiler-driven software with transformer-optimized hardware, and its LISA™ (Language Instruction Set Architecture) scales a single toolchain from 4-billion-parameter models to the trillion-parameter class. The HTX301-4S card has an approximately 155W card-level power figure; workload-specific and whole-system power remain subject to engineering validation. Multi-card systems are designed to support larger model-memory requirements while enabling enterprises to retain control of models, data, and deployment. Learn more at skymizer.ai.

HyperThought™ and LISA™ are trademarks of Skymizer Taiwan Inc. All other trademarks are the property of their respective owners.