Mistral released Voxtral, its first open-weight family for transcription and audio understanding. The official benchmarks show strong multilingual speech recognition and audio Q&A, while the two model sizes have very different hardware requirements. Here is what the release actually supports and why Whisper Notes still uses smaller, transcription-focused models on consumer Apple hardware.
Two Models
Mistral released two Apache-2.0 checkpoints in July 2025:
Voxtral Small 24B
- •24 billion parameters
- •Transcription, translation, audio Q&A, and summarization
- •32k context window
- •About 55 GB of GPU RAM in bf16/fp16
Voxtral Mini 3B
- •3 billion parameters
- •Same audio-and-text task family in a smaller checkpoint
- •Officially positioned for local and edge deployment
- •About 9.5 GB of GPU RAM in bf16/fp16
Open Weights and License
Both 2025 checkpoints are open weights under the Apache 2.0 license. You can download and run them yourself:
- ✓ Full model weights available
- ✓ Self-host or adapt them within the Apache 2.0 license
- ✓ No Mistral API fee when self-hosting; you still pay for your own compute
- ✓ Process audio on your own servers
Sources: Mistral's Voxtral release, the official Voxtral Small model card, and the official Voxtral Mini model card.
Benchmarks
Mistral's release compares macro-average word error rate across English short-form and long-form sets, Mozilla Common Voice, FLEURS, and Multilingual LibriSpeech. In Mistral's reported FLEURS results, Voxtral Small beats Whisper on every evaluated task. Lower WER is better:
FLEURS WER from Mistral's launch benchmarks plotted against estimated API price; both benchmark results and prices can change
FLEURS is only one slice of transcription quality. These are vendor-reported results, so compare the published benchmark definitions and test your own languages, microphones, and recording conditions before choosing a model.
Voxtral vs Whisper: Which Should You Use?
Benchmarks only tell part of the story. Here's how the two model families compare on the factors that matter if you actually want to run one yourself:
| Voxtral Small | Voxtral Mini | Whisper Large v3 | Whisper Large-v3 Turbo | |
|---|---|---|---|---|
| Languages | 8 primary languages | 8 primary languages | 100+ | 100+ |
| Open source | Yes | Yes | Yes | Yes |
| Model size | 24B parameters; ~55 GB GPU RAM in bf16/fp16 | 3B parameters; ~9.5 GB GPU RAM in bf16/fp16 | 1.55B parameters | 809M parameters (~1.6 GB) |
| Practical on a consumer Mac | Not in full precision on typical hardware | Possible with a compatible runtime; heavier than Turbo | Yes | Yes |
The verdict: Voxtral Small reports strong WER results and adds audio understanding that Whisper does not attempt. Voxtral Mini is the checkpoint Mistral positions for local and edge deployment, but its official model card still lists about 9.5 GB of GPU RAM in bf16/fp16. For a consumer transcription app, Whisper Large-v3 Turbo remains easier to ship and covers far more languages. That is why Whisper Notes currently combines Whisper Turbo with Parakeet V3 and SenseVoice rather than Voxtral.
API and Self-Hosting Cost
As checked on July 10, 2026, Mistral's model card lists Voxtral Small audio input at $0.004 per minute. Text input and generated text are billed separately for audio-understanding requests. If you self-host the Apache-2.0 weights, there is no Mistral API charge, but you pay for the hardware and operations.
Mistral API
Self-hosted weights
How It Works
The official model card describes Voxtral as a language-model backbone with an audio encoder and dedicated transcription mode. The important product limits are concrete:
1. Multimodal Architecture
Voxtral accepts audio and text in the same conversation instead of acting only as a speech-to-text decoder:
- •Dedicated mode for direct transcription
- •Audio Q&A and structured summarization
- •Multiple audio inputs and multi-turn audio conversations
Long-form limit
The 32k context window supports up to 30 minutes of audio for transcription or 40 minutes for broader audio understanding.
2. Languages and Evaluation
Mistral lists eight primary languages with automatic detection:
- •English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian
- •Benchmarks include FLEURS, Mozilla Common Voice, and Multilingual LibriSpeech
- •The release also evaluates English short-form and long-form audio separately
3. Deployment Requirements
Open weights do not mean every checkpoint is small enough for every device:
- •Voxtral Small requires about 55 GB of GPU RAM in bf16/fp16 according to its model card
- •Voxtral Mini is the 3B checkpoint intended for local and edge deployment
- •Mistral recommends vLLM for serving the released checkpoints
4. Key Features
Contextual Understanding
Can answer questions and produce summaries directly from audio, within the 32k context limit.
Multilingual
Automatically detects and transcribes eight primary languages: English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian.
Noise Handling
The official evaluation includes varied speech datasets, but Mistral does not publish a blanket guarantee for every noisy recording.
Edge Deployment
Voxtral Mini is the local and edge option. Its official model card lists about 9.5 GB of GPU RAM in bf16/fp16.
5. Architecture
Three main components:
- 1. Audio encoder: Converts the waveform into learned audio representations
- 2. Projector: Maps audio representations into the language model's hidden space
- 3. Language-model backbone: Generates transcripts, answers, summaries, or tool calls
That design explains why Voxtral can do more than transcription, but it also carries much larger language-model weights than a transcription-only model such as Whisper Turbo.
Why Whisper Notes Still Makes Sense
Voxtral and Whisper Notes solve different problems. Voxtral combines transcription with audio understanding; Whisper Notes focuses on private transcription that is easy to run on consumer Apple hardware.
What Whisper Notes Offers
Privacy
- •100% offline processing
- •No data transmission
- •No cloud dependencies
Performance
- •Whisper technology, proven accuracy
- •Optimized for Apple Silicon
- •Reliable results
Cost
- •One-time pricing; App Store bundle and direct-download DMG sold separately
- •No per-minute charges
- •Unlimited transcription
User Experience
- •Simple interface
- •Regular updates
- •Continuous improvements
Storage Requirements
Voxtral Small is a server-class model in full precision: its official card lists about 55 GB of GPU RAM. Voxtral Mini is specifically intended for local and edge use, but its card still lists about 9.5 GB of GPU RAM in bf16/fp16. Whisper Turbo's app download is roughly 1.6 GB, which is a better fit for the current Whisper Notes product.
Whisper Notes uses Whisper Large-v3 Turbo—it balances performance, speed, and VRAM requirements for everyday use. We'll upgrade to better models when they become available with reasonable resource requirements.
Voxtral is attractive when audio Q&A, summarization, or self-hosted server deployment matters. Whisper Notes is aimed at individual users who want private transcription and no recurring subscription.
What This Means
Voxtral is an important open-weight step beyond plain speech recognition because it can reason over audio as well as transcribe it.
For private offline transcription on Mac and iPhone, smaller specialized engines remain easier to package and run consistently.
Comparing all the local options? Every on-device speech-to-text model — Whisper variants, Parakeet V3, SenseVoice, and Voxtral — is compared side by side on our Whisper models comparison page.
Frequently Asked Questions
What is Mistral Voxtral?
Voxtral is Mistral's open-weight family for speech transcription and audio understanding. The July 2025 release includes Voxtral Small 24B for server-class workloads and Voxtral Mini 3B for local and edge deployment. Both checkpoints use the Apache 2.0 license.
Can Voxtral run locally on a Mac?
The downloadable weights can be self-hosted, but the two sizes have different requirements. Voxtral Small needs about 55 GB of GPU RAM in bf16/fp16, so it is a server-class choice. Voxtral Mini is the local and edge checkpoint; its model card lists about 9.5 GB in bf16/fp16, and actual Mac speed and memory use depend on the runtime and quantization.
Does Whisper Notes use Voxtral?
No. Whisper Notes runs Parakeet V3 (the Mac default), Whisper Large-v3 Turbo, and SenseVoice — models chosen because they balance performance, speed, and VRAM requirements for everyday use on Apple Silicon. We'll adopt better models whenever they become practical to run on consumer hardware.
How many languages does Voxtral support?
The official model cards list eight primary languages with automatic detection: English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian. Mistral also evaluates Arabic in FLEURS, but it is not included in the model card's primary-language list.
When was Mistral Voxtral released?
Mistral released Voxtral on July 15, 2025. The initial Apache-2.0 checkpoints were Voxtral Small 24B and Voxtral Mini 3B.
Download for iOS