Voxtral Small 24B (2507)

Architecture: Multimodal audio-language model pairing a Whisper-style audio encoder with the Mistral Small 3.1 (24B) decoder LLM via a multi-modal projector. Implemented as VoxtralForConditionalGeneration in Hugging Face Transformers; recommended deployment via vLLM with --tensor-parallel-size 2 (≈55 GB GPU RAM in bf16/fp16).

Modality: Audio + text in / text out. 32k-token context window, handling audios up to ~30 min for transcription and ~40 min for audio understanding.

Capabilities:

• State-of-the-art transcription that beats Whisper large-v3 across all tested FLEURS languages and matches/exceeds GPT-4o mini Transcribe, Gemini 2.5 Flash and ElevenLabs Scribe on Mozilla Common Voice and English short-form.
• Audio Q&A and structured summarization end-to-end without chaining a separate ASR.
• Speech translation — state-of-the-art on FLEURS-Translation.
• Function calling from voice with --tool-call-parser mistral --enable-auto-tool-choice.
• Natively multilingual: English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian (and more), with automatic language detection.
• Retains the full text intelligence of Mistral Small 3.1 (matches its MMLU/DROP curves).

Training & release: Released July 15, 2025 alongside Voxtral Mini (paper arXiv 2507.13264). Distributed under Apache 2.0; available on Mistral's La Plateforme, Le Chat voice mode, and via private deployment.

Use cases: Production-scale transcription, multilingual voice agents, meeting/call analysis, podcast and media indexing, voice-driven workflow automation in regulated environments.