MetaX's 300TOPS edge terminal beats NVIDIA Thor T5000 in tests, trains robot duck locally

机器之心 · wechat · 2026-11-04

Chinese GPU startup MetaX (Days Intelligent) gave its first full media showcase since its January Hong Kong IPO, headlined by the Tongyang TY1200 edge AI terminal.

Highlights:

Replicated Jensen Huang's GTC Microduck robot duck demo, but trained entirely on-device: one TY1200 (300TOPS dense compute) handled environment building, policy training, and real-time simulation locally — claimed as the first local GPU-accelerated physics simulation on an edge device in China
Benchmarks vs. NVIDIA Thor T5000: 1.38x throughput on Qwen3.8-27B-AWQ, 1.80x on Qwen2.5-VL-3B-Instruct, up to 1.80x on Cosmos-R2 VLA models; first token in 0.223s vs. 0.411s, 30.21 vs. 24.46 tok/s
Deployments: battery-material analysis at Hohai University (2 min reports), robot dog with fully local face detection, museum guide that trains CV models overnight, and a guide robot for visually impaired patients in Shanghai
"AI to develop AI": FlyDSL + agentic workflows compress 10-person-day kernel development to 1 person-hour; PD-separated deployment under 10 minutes; typical optimization from 7 days to 2 hours
Roadmap: a 5-product edge matrix with cloud-shared software stack, plus a 600TOPS TY2000 preview

The article cites Stanford HAI data — the smallest model crossing 60% on MMLU shrank from 540B params (2022) to 3.8B (2024) — as the driver behind edge-side "compute layering."

Original post →

MetaX's 300TOPS edge terminal beats NVIDIA Thor T5000 in tests, trains robot duck locally

More from Embodied