Update: MiMo-V2.6-Pro UltraSpeed — 10x the speed, 10x the price
Xiaomi now serves its just-open-sourced flagship through a closed fast stack; the base MIT weights on Hugging Face are still your self-host escape hatch.
Copy markdown
10x the throughput, billed at exactly 10x
MiMo-V2.6-Pro-UltraSpeed went live on OpenRouter Sept 21 — the same trillion-parameter checkpoint as the base model, tuned for speed at $4.35/$8.70 per million tokens, precisely 10x the base $0.435/$0.87. Xiaomi claims roughly 10x output throughput, but early third-party numbers (OpenRouter's own listing shows ~125 tok/s) don't yet show that gap over the base tier, so benchmark your own workload before paying the premium.
The catch: the fast tier is closed and cache-free
UltraSpeed ships no weights — only the serving stack is proprietary — while the base MiMo-V2.6-Pro keeps its MIT license on Hugging Face, so self-hosting stays your fallback at the 1x price. UltraSpeed also lists no cache-discount tier, so cache-heavy agent loops that lean on prompt reuse lose the savings and pay full rate on every token.