Hacker News front pageJun 8, 2026, 23:27 UTC+8

MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second

The title says Xiaomi MiMo-v2.5-Pro-UltraSpeed is a 1T model running at 1,000 tokens per second; the RSS body only provides the URL, Hacker News comments link, 66 points, and 14 comments, and the post does not disclose hardware, precision, context window, benchmark setup, or availability.

Why it matters: HKR-H/K/R all pass: Xiaomi’s MiMo update has a sharp 1T/1,000 tokens/s claim and clear cost-speed resonance. Missing hardware, precision, context window, and test setup keep it in the 78–84 band, not p1.

Read the original ↗Export Markdown

Take

Xiaomi is showing inference engineering, not model intelligence. MiMo-V2.5-Pro-UltraSpeed claims a 1T model at 1,000+ tokens/s, with the demo saying up to about 1,200 tokens/s. The machinery is explicit: FP4 quantization, DFlash speculative decoding, TileRT co-design, and an application-only API from June 9 to June 23. That setup says the speed comes from a tightly controlled model-system stack, not from cheap general serving at scale.

The pricing tells the same story: 3× MiMo-V2.5-Pro for roughly 10× generation speed, but Xiaomi gives no base token price, GPU type, batch size, context window, or quality regression after FP4. The Cerebras/Groq comparison is clever because Xiaomi wants credit for doing this on commodity GPUs. Fine — but without quality and serving conditions, 1,000 tps proves a serious low-latency demo, not a production-grade flagship model.

Source: Hacker News front page · mimo.xiaomi.com
Published: Jun 8, 2026, 23:27 UTC+8 · Jun 8, 2026