{
  "id": 18462175,
  "title": "Skymizer and MiTAC run 2 8 trillion parameter AI model on server",
  "url": "https://urgent.news/2026/11/06/skymizer-and-mitac-run-2-8-trillion-parameter-ai-model-on-server",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-06T07:38:24.000Z",
  "source": {
    "name": "Vietnam Investment Review",
    "slug": "vietnam-investment-review",
    "url": "https://vir.com.vn/skymizer-and-mitac-run-28-trillion-parameter-ai-model-on-server-164587.html"
  },
  "original_language": "en",
  "account": "TAIPEI, Nov. 5, 2026 - Skymizer and MiTAC Computing jointly unveiled Moonshot AI's 2.8-trillion-parameter Kimi K3 model at SC26, running on a MiTAC server with eight Skymizer HTX301 LPU cards and 2 TB of card-attached memory. The 2 TB memory allows for the model's 1.56 TB weight set and KV cache. Skymizer will showcase the technology at ICCAD-Expo 2026 in Beijing. The joint effort builds on Skymizer's Qwen3-VL-30B-A3B model demonstrated at COMPUTEX 2026. Skymizer's LISA™ platform provides common instruction set and toolchain for memory-residency and scheduling of larger models. Enterprises seek control over AI models and data, and Skymizer's work with MiTAC aims to make open-weight models more practical for on-premises deployment. The collaboration explores Prefill–Decode disaggregation and optimized AI inference deployment. Skymizer is developing an AI-assisted delivery approach to help enterprises select models, prepare environments, install stacks, and validate deployments while retaining enterprise control over models, data, and AI operations. HTX301 chip production is anticipated to begin in Q1 2027.",
  "summary": "Technology vendors Skymizer and MiTAC Computing will showcase the 2 8 trillion parameter Kimi K3 artificial intelligence model running on a single air cooled server configured with two terabytes of card memory",
  "key_points": [
    "Skymizer and MiTAC jointly unveil 2.8-trillion-parameter Kimi K3 model",
    "Moonshot AI model runs on MiTAC server with eight HTX301 LPU cards",
    "Collaboration aims to make open-weight models practical for on-premises deployment"
  ],
  "editors_take": "This collaboration provides enterprises with the hardware and software infrastructure necessary to deploy massive open-weight models on-premises while maintaining strict control over their proprietary data and AI operations.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}