{
  "id": 16242620,
  "title": "Plucky Chinese startup claims to beat Nvidia DGX Spark, Mac Mini on Qwen — says its '256 TOPS beats 1000 TOPS in prefill', running 16-core Cortex-A55 SoC on 40w",
  "url": "https://urgent.news/2026/10/23/plucky-chinese-startup-claims-to-beat-nvidia-dgx-spark-mac-mini-on",
  "topic": "culture",
  "section": "Culture",
  "published": "2026-10-23T22:25:00.000Z",
  "source": {
    "name": "TechRadar",
    "slug": "techradar",
    "url": "https://www.techradar.com/pro/plucky-chinese-startup-claims-to-beat-nvidia-dgx-spark-mac-mini-on-qwen-says-its-256-tops-beats-1000-tops-in-prefill-running-16-core-cortex-a55-soc-on-40w"
  },
  "original_language": "en",
  "account": "China's Sipeed Electronics has announced its latest AI mini PC, the MaixStation, which it claims outperforms Nvidia's DGX Spark in certain tasks. The 128GB model costs less than $4,000, while Nvidia's DGX Spark is priced at $6,949.99. Both machines boast impressive specifications, including 16 Arm Cortex-A55 cores and 256 TOPS for AI acceleration, but Sipeed emphasizes its more efficient processing during prefill operations. Nvidia advertises 1,000 TOPS using sparse FP4 calculations, while Sipeed argues that practical workloads cannot fully exploit this theoretical capacity. The compact machine features 64 video decoding channels, 16 encoding channels, and supports memory configurations of up to 128GB. Price and availability remain tentative, with Sipeed planning a Kickstarter campaign within months and offering a $100 discount for early reservations.",
  "summary": "Sipeed claims its 256 TOPS MaixStation can outperform Nvidia's DGX Spark on Qwen workloads despite lower advertised computing capacity and consumption.",
  "key_points": [],
  "editors_take": null,
  "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."
}