Microsoft AI: MAI-Transcribe 2

Microsoft AI: MAI-Transcribe 2 is a speech-to-text model from Microsoft, released on 3 September 2026. It sits in the transcription category, so its job is narrow and specific: turning spoken audio into written text. Microsoft positions it as a multilingual system, and it is ranked first on the FLEURS multilingual benchmark according to the provider. If your work involves speech in more than one language, that combination of provider backing and benchmark placement is the main reason to look at this model rather than a general purpose one.

The model takes audio as input and returns a transcription as output. It covers 60 languages and performs automatic language identification, so you do not have to tell it which language is being spoken before sending a request. It also handles code switching, meaning audio where speakers move between languages within the same recording. The published details do not list a context length in tokens, and the maximum output is given as zero, so treat the model as a pure transcription endpoint rather than something you can prompt with long text instructions. There are no supported request parameters listed beyond what the API requires to identify the model and pass audio.

The practical uses follow directly from those capabilities. Meeting and call recordings in multilingual organisations are an obvious fit, particularly where participants switch languages mid conversation. Media and podcast workflows can use it to produce transcripts without a separate language detection step. Research and archival projects dealing with speech in any of the 60 supported languages can send audio through and get text back. Because language identification is automatic, batch jobs over mixed-language audio do not need to be sorted or tagged beforehand.

You call the model through api14's OpenAI-compatible API using the model id "microsoft/mai-transcribe-2". Because the interface follows the OpenAI format, existing clients and SDKs that already point at an OpenAI-compatible base URL can be reused, with the audio sent as the input and the transcription returned in the response. The model lists no supported parameters, so keep requests minimal: set the model id, attach the audio, and avoid assuming that sampling or formatting options will have any effect. If a parameter is not listed as supported, it is safest to leave it out.

On limits, the available facts are thin. There is no stated context length and the maximum output is zero, which suggests the model is not intended for text-in, text-out prompting or for long generated responses. If your task needs reasoning over a transcript, summarisation, translation, or any text generation, a text model is the better choice, with MAI-Transcribe 2 used first to produce the transcript. If your audio falls outside the 60 supported languages, this model is unlikely to help.