Mistral drops Voxtral to challenge Whisper with open speech AI

Voxtral is Mistral AI's new family of open-source speech understanding models, released under the Apache 2.0 license in 24B and 3B parameter sizes. Instead of merely transcribing audio, it natively fuses speech recognition and language reasoning into a single model, letting you ask questions about recordings, generate structured summaries, and trigger system actions straight from spoken words.

What it can do

Building voice applications usually means chaining two separate systems: an ASR model to transcribe speech into text, and an LLM to figure out what that text means. That chain adds latency and drops subtle audio context. Voxtral merges both jobs into one pass.

With a 32k token context window, Voxtral handles up to 30 minutes of audio for transcription and up to 40 minutes for deep understanding. You can drop in a full podcast or meeting recording and ask specific questions without running an external text model.

Even better, it supports direct function calling straight from voice. A user speaks a command, and the model triggers backend APIs or workflows without needing intermediate text parsing.

Voxtral is natively multilingual, handling languages like English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian with automatic language detection. Because it is built on the Mistral Small 3.1 backbone, it retains full text capabilities and can serve as a drop-in replacement for text workloads.

Specs and how to run it

Mistral released two variants. Voxtral Small has 24 billion parameters and is aimed at production-scale inference. Voxtral Mini sits at 3 billion parameters, tailored for edge devices and local desktop setups.

On transcription benchmarks, Voxtral comprehensively outperforms Whisper large-v3, while beating GPT-4o mini Transcribe and Gemini 2.5 Flash across multiple tasks. According to Mistral, Voxtral Small matches ElevenLabs Scribe at less than half the price.

Both the 24B and 3B weights are free to download on Hugging Face under Apache 2.0. If you prefer cloud endpoints, Mistral offers an API starting at $0.001 per minute, featuring an optimized Voxtral Mini Transcribe route for raw transcription speed. Voice mode inside Mistral's Le Chat is also rolling out over the next few weeks.

Why it matters

If you have been relying on Whisper for speech pipelines or avoiding closed voice APIs because of steep billing and privacy concerns, Voxtral is the first open drop-in replacement you should test right now. Combining audio perception and reasoning into one open model removes pipeline complexity and slashes inference costs.

Keep in mind that it still lacks granular features like speaker diarization, word-level timestamps, and emotion tagging, which Mistral says are coming down the road. But for voice agents that need to listen and take real backend action immediately, open-source audio just caught up with proprietary platforms.