Voxtral

weights

open(Voxtral Mini 4B Realtime and Voxtral Mini 3B as safetensors on the Hub, ungated)

data

closed(the cards describe no training data)

code

partial(inference through vLLM, Transformers and mistral-common

license

Apache-2.0(the speech-recognition line, Voxtral Mini 4B Realtime 2602 and Voxtral Mini 3B 2507)

separate-line

Voxtral 4B TTS 2603, CC-BY-NC-4.0(the text-to-speech model, announced separately with its own name, card and version, and licensed CC-BY-NC-4.0 because it inherits the license of its bundled reference voices

api-only-tier

Voxtral Mini Transcribe 2.0(the offline transcription model the Realtime card compares against

Mistral's current speech-recognition models, Voxtral Mini 4B Realtime and Mini 3B, download freely under Apache 2.0. Voxtral 4B TTS is a separate line, announced on its own with its own card, and licensed CC-BY-NC-4.0 because of the reference voices it ships with, so it may not be used commercially; its terms do not carry over to the recognition models. The cards describe no training data and only inference code is published.

• https://huggingface.co/api/models?author=mistralai&search=Voxtral&limit=100&expand[]=downloads&expand[]=createdAt&expand[]=tags&expand[]=gated recorded 2026-09-27

  Family listing: Voxtral-Mini-4B-Realtime-2602, Voxtral-Mini-3B-2507 and Voxtral-Small-24B-2507 tagged "license:apache-2.0"; Voxtral-4B-TTS-2603 tagged "license:cc-by-nc-4.0"; all ungated.

• https://huggingface.co/api/models/mistralai/Voxtral-4B-TTS-2603 recorded 2026-09-27

  Hub record: "gated":false, with consolidated.safetensors and voice_embedding files in the file list.

• https://huggingface.co/api/models/mistralai/Voxtral-Mini-3B-2507 recorded 2026-09-27

  Hub record: "gated":false, with consolidated.safetensors and sharded safetensors in the file list.

• https://huggingface.co/api/models/mistralai/Voxtral-Mini-4B-Realtime-2602 recorded 2026-09-27

  Hub record: "gated":false, with model.safetensors in the file list.

• https://huggingface.co/mistralai/Voxtral-4B-TTS-2603/raw/main/README.md recorded 2026-09-27

  Card: "These voices are licensed under CC BY-NC 4, which is the license that the model inherits."; "Thus, this model inherits the same license."

• https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602/raw/main/README.md recorded 2026-09-27

  Card: "This model is released in **BF16** under the **Apache-2 license**, ensuring flexibility for both research and commercial use."; the comparison table lists "Voxtral Mini Transcribe 2.0" as an offline model.

• https://mistral.ai/news/voxtral-transcribe-2 recorded 2026-09-28

  "Voxtral transcribes at the speed of sound", the recognition line's announcement: "The family includes Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live applications". It does not mention the TTS model.

• https://mistral.ai/news/voxtral-tts recorded 2026-09-28

  "Speaking of Voxtral", Mistral's own announcement of the TTS model: "A model with several reference voices is available as open weights on Hugging Face under CC BY NC 4.0 license", and "It works alongside Voxtral Transcribe for full speech-to-speech, or integrates into any existing speech-to-text and LLM stack".

Hugging Face downloads summed over the three declared checkpoints, almost all of them the streaming recognizer.

• https://huggingface.co/api/models?author=mistralai&search=Voxtral&limit=100&expand[]=downloads&expand[]=createdAt&expand[]=tags&expand[]=gated recorded 2026-09-27

  Trailing-30-day downloads: Voxtral-Mini-4B-Realtime-2602 1,765,082, Voxtral-Mini-3B-2507 247,167, Voxtral-4B-TTS-2603 1,583.

Mid-table on both boards, level with Whisper for recognition and close to Kokoro for synthesis; the realtime model trades some accuracy for streaming with a sub-second delay.

• https://artificialanalysis.ai/text-to-speech/leaderboard recorded 2026-09-27

  Leaderboard entry "name":"Voxtral TTS" (creator Mistral) with "openWeights":true, ranked 42nd of the 91 models on the board.

• https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/d2c5b384deccdb82834f41aeaffcc618c00efa2f/english_short_latest.csv recorded 2026-09-27

  Results CSV rows "mistralai/Voxtral-Mini-3B-2507,5.53875" and "mistralai/Voxtral-Mini-4B-Realtime-2602,6.4625".

• https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602/raw/main/README.md recorded 2026-09-27

  Card: "Built with a **natively streaming architecture** and a custom causal audio encoder - it allows configurable transcription delays (240ms to 2.4s)".