Skymizer Taiwan Inc. and MiTAC Computing Technology Corp. plan a single-server demonstration of Moonshot AI’s 2.8-trillion-parameter Kimi K3 at SC26, with system validation still in progress.
The planned system pairs eight HTX301 language-processing-unit (LPU) cards with 2 TB of card-attached memory, Skymizer said. The company estimates Kimi K3’s weights at about 1.56 TB and plans to reserve some capacity for a key-value (KV) cache. Engineers are still determining memory allocation and which workloads the server can support.
Moonshot AI’s model card describes Kimi K3 as a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters. It says the model selects 16 of 896 routed experts for each token and lists MXFP4 weights and MXFP8 activations.
Skymizer characterizes HTX301 as a decode-first design focused on token generation. Its product page separately describes a high-capacity accelerator configuration as supporting up to about 1 trillion parameters, but does not explain how that figure maps to the planned eight-card Kimi K3 system. The announcement provides no system-level cost, throughput or latency figures.
Skymizer identifies financial services, healthcare, government, manufacturing and semiconductor design as potential markets for on-premises systems. It says local deployment may appeal to organizations seeking closer oversight of their data and AI operations. The announcement lists about 155 watts for each HTX301-4S card; workload-specific and whole-server power still require engineering validation. MiTAC said the collaboration is also exploring separate handling for prompt processing (prefill) and token generation (decode) across AI workloads.
Skymizer expects chip production to begin in Q1 2027 and says server development and validation are continuing. The company says sampling and broader commercial availability will be addressed separately. SC26 is scheduled for Nov. 15-20 in Chicago, where the demonstration is planned.
