Xiaomi · updated 2026-08-29
Run MiMo V2.5: providers, pricing and flat-rate access
Xiaomi's April 2026 open-weight (MIT) omnimodal model — a 310B-parameter sparse MoE with 15B active, handling text, image, audio and video, positioned as frontier-level agentic capability at commodity prices.
MiMo V2.5 is very cheap to run: Xiaomi's first-party API charges $0.14/$0.28 per M tokens with deep cache discounts, and the MIT open weights mean six-plus competing hosts (DeepInfra is fastest at 130+ tokens/sec). Light users should pay per token. Standard Compute serves both V2.5 and V2.5 Pro in its flat-rate routed pool.
Your options, honestly
Xiaomi APIPay-per-tokenFirst-party and cheapest — $0.14/$0.28 per M tokens with unusually aggressive prompt-caching discounts.
DeepInfra / OpenRouter hostsPay-per-tokenSix-plus providers on the open weights; DeepInfra leads on speed (130+ t/s). Price and latency vary by host.
Standard Compute$39+/mo flatV2.5 and V2.5 Pro in the routed pool — omnimodal workhorse steps at a fixed price, frontier models behind them for hard reasoning. No pinning.
Self-host (vLLM)GPU + powerMIT-licensed weights on Hugging Face; the 15B-active MoE is more hostable than its 310B total suggests, but it's still serious hardware.
the flat-rate trade-off
At $0.14/M input, a MiMo-only workload should honestly just pay per token. The flat-rate case is the usual one: MiMo-class volume (especially audio/video understanding, where it's one of few cheap options) plus frontier-class hard steps in one agent, under one fixed price instead of two unpredictable bills.
Running it in your agent
OpenAI-compatible agents run MiMo V2.5 with a base-URL swap to any host serving it — api.stdcmpt.com/v1 for flat-rate. On Standard Compute the router sends suitable routine and multimodal steps to MiMo-class automatically.
FAQ
MiMo V2.5 or V2.5 Pro?
Pro is the stronger text-reasoning variant at roughly 4x the base model's price — still cheap in absolute terms. Base V2.5 is the omnimodal workhorse. Routed plans serve both and make the split per-request.
Can Xiaomi's model really handle agentic coding?
It's credible for routine agent steps and stands out for omnimodal input (audio/video few cheap models handle), but frontier models still lead on hard multi-file reasoning. Treat it as the volume tier in a mixed stack, not a frontier replacement.
More: how flat-rate LLM APIs work · per-agent guides for OpenCode, Claude Code and Cline · model comparisons