News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
wavVoxtral

Voxtral

Open audio chat models that transcribe, answer questions and summarise speech

Voxtral is a pair of multimodal audio chat models from Mistral AI, released in July 2025 in a 24B and a 3B size under Apache 2.0. Built on the Mistral Small 3.1 language backbone, they transcribe and understand speech in many languages while keeping strong text skills, and can be run locally or called through an API.

Share on WhatsApp

  • Two sizes: 24B Voxtral Small and 3B Voxtral Mini, both Apache 2.0
  • 32k context handles audio up to 30 minutes for transcription, 40 for understanding
  • Transcribes, answers questions, summarises and calls functions from speech
  • Built on the Mistral Small 3.1 language backbone
  • API pricing starts at $0.001 per minute

What it is

Voxtral is a family of two speech understanding models from Mistral AI: Voxtral Small at 24 billion parameters for production use, and Voxtral Mini at 3 billion for local and edge deployment. They take in spoken audio and text, and can transcribe, answer questions about recordings, produce summaries, translate speech and trigger functions from spoken requests. Both sizes were put out under the Apache 2.0 licence in July 2025.

How it came about

Mistral AI announced Voxtral in July 2025 alongside a research paper by Alexander H. Liu and more than a hundred co-authors, posted to arXiv on 17 July 2025. The company frames the release as a move towards voice as a natural way to interact with computers, and says its audio team is still small and growing. A webinar with Inworld on voice agents was scheduled for 6 August 2025.

Inside the model

The two models share a 32k token context window, which Mistral AI says covers audio files up to 30 minutes for transcription and 40 minutes for understanding. The language backbone is Mistral Small 3.1, and the company says text skills survive the audio training. Voxtral does transcription, question answering and summarisation in one model rather than chaining a speech recogniser to a language model. It detects the language of the audio automatically and handles function calling straight from speech. A transcribe-optimised variant, Voxtral Mini Transcribe, backs the API's transcription route.

What it can do

Mistral AI reports macro-average word error rates on English and multilingual sets, saying Voxtral beats Whisper large-v3 across the board and also beats GPT-4o mini Transcribe and Gemini 2.5 Flash on every task, with top results on English short-form and Mozilla Common Voice. On FLEURS, Voxtral Small is reported ahead of Whisper on every task. For audio understanding, the company says Voxtral Small is competitive with GPT-4o-mini and Gemini 2.5 Flash and leads on speech translation. The paper also contributes three benchmarks for knowledge and trivia in speech.

How to use it

Both sizes are downloadable from Hugging Face under Apache 2.0. Mistral AI also offers an API starting at $0.001 per minute, and was rolling Voxtral into the voice mode of Le Chat on web and mobile. Enterprise options include private deployment, domain fine-tuning and integration support.

Safety and testing

Mistral AI says it is inviting design partners to help build speaker identification, emotion detection, advanced diarisation and longer context windows. Age and emotion markup in audio is listed as forthcoming work.

Reception and debate

Mistral AI positions Voxtral against open speech recognition systems it describes as having high word error rates, and against closed APIs it describes as costly and hard to control. The company claims Voxtral Mini Transcribe beats OpenAI Whisper and Voxtral Small matches ElevenLabs Scribe, both at less than half the price.

Why it matters

Mistral AI describes the release as a step towards natural voice interfaces and says more audio work is planned. The models are offered as drop-in replacements for Ministral and Mistral Small 3.1 in text tasks, and the company is recruiting for its audio team.

Questions people ask about Voxtral

What is Voxtral?

It is a pair of audio chat models from Mistral AI, in 24B and 3B sizes, released in July 2025. They transcribe speech, answer questions about recordings, summarise audio and can trigger functions from spoken requests, while keeping the text skills of their Mistral Small 3.1 backbone.

How can I use it?

Both sizes are on Hugging Face under Apache 2.0, so you can run them yourself. Mistral AI also offers an API from $0.001 per minute and was adding Voxtral to the voice mode of Le Chat on web and mobile.

What can it handle?

A 32k token context window lets it work on audio up to 30 minutes for transcription and 40 minutes for understanding. It detects language automatically and covers widely used languages including English, Spanish, French, Portuguese, Hindi, German, Dutch and Italian.

How does Mistral AI say it compares?

The company reports lower word error rates than Whisper large-v3 across its tests, and says it beats GPT-4o mini Transcribe and Gemini 2.5 Flash on every task. It claims Voxtral Small matches ElevenLabs Scribe at less than half the price.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Voxtral yet, so the facts on this page come from the sources listed above: mistral.ai/news/voxtral, arxiv.org/abs/2507.13264. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.