Voxtral
2025-07-15 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI launches Voxtral, its first speech understanding models: Small (24B) and Mini (3B), under Apache 2.0, to transcribe and query audio, starting at $0.001 per minute.
What changes
- Two sizes: 24B for production, 3B for local and edge use; the Mini Transcribe version is optimized for the transcription API.
- 32k-token context: audio up to 30 minutes for transcription, 40 minutes for understanding.
- Questions and answers and summaries directly from audio, function calls triggered by voice.
- Automatic language detection; claimed to be better than Whisper large-v3 and close to ElevenLabs Scribe for less than half the price.
### Specifications - Based on Mistral Small 3.1 (Small) and Ministral (Mini), which retain text capabilities. - Inputs: audio and text; text output. - License: Apache 2.0. - Availability: API, Hugging Face, Le Chat voice mode. - API identifiers: `voxtral-small-2507`, `voxtral-mini-2507` (chat), `voxtral-mini-2507` via audio/transcriptions. ### Pricing | Model | Audio (USD / min) | Text input (USD / M tokens) | Text output (USD / M tokens) | |---|---|---|---| | Voxtral Mini (transcription) | 0.002 | | | | Voxtral Mini | 0.001 | 0.04 | 0.04 | | Voxtral Small | 0.004 | 0.1 | 0.3 | Prices from the pricing page archived on July 30, 2025; the current documentation shows 0.4 USD for text output for Voxtral Small. The post states "starting at 0.001 USD per minute". ### Benchmarks The results (word error rate on FLEURS, Common Voice, etc.) are provided as image charts, with no figures in text. The vendor states that Voxtral outperforms Whisper large-v3, GPT-4o mini Transcribe and Gemini 2.5 Flash on its transcription tasks. ### Key takeaways An open speech building block, deployable on-premises, that avoids chaining separate speech recognition and LLMs.
Links
Tags : LLM, Mistral, Voxtral, Audio, Open Source, Apache 2.0