MiMo-V2.5 Voice

MiMo-V2.5-ASR fills a specific niche: Chinese-English bilingual and dialect transcription at a very low price. If you're building an app that must handle Mandarin, Cantonese, or eight other Chinese dialects, this is one of the cheapest per-hour options around. The $0.074/hour rate is dramatically lower than many general-purpose ASR APIs, which often charge per minute or per request with higher effective costs. For a media company captioning thousands of hours of Cantonese and Mandarin content, that per-hour price is the headline feature. The open 8B weights are another differentiator. Teams that need on-prem or offline transcription can self-host, which is rare among commercial ASR APIs. Combined with OpenAI- and Anthropic-compatible APIs, you can slot it into existing pipelines without vendor lock-in. The V2.5 series also means you're not on a deprecated model; V2 was retired June 30, 2026, and Xiaomi is pushing migration. Where it falls short: no real-time streaming and no speaker diarization. If you need live captioning for a meeting or want to know who said what in a call-center recording, this isn't for you. It's also focused on Chinese and English, so other languages won't work well. And despite the low per-hour rate, costs scale linearly—transcribe 10,000 hours a month and you're paying $740, which may still beat alternatives but warrants monitoring. The news signal is thin—Xiaomi doesn't document a public changelog with dates for the ASR model itself. What's confirmed is that this is a solid, specialized choice for high-volume bilingual transcription, not a general ASR solution. If your need is narrow and you value low cost + open weights, it's a smart pick. If you need streaming or speaker labels, look at services like Deepgram or AssemblyAI, even though they cost more per hour.