Voxtral Small 24B 2507 (Amazon Bedrock, US)

Voxtral Small 24B 2507 is a multimodal model that builds on Mistral Small 3, adding audio input handling while preserving strong language understanding. The 24-billion parameter design targets speech transcription, translation, and audio comprehension alongside conventional text tasks, so a single model can process spoken and written inputs without routing to a separate speech engine. Built-in question answering and summarization let users query audio directly and produce structured outputs, while a dedicated transcription mode is tuned for pure speech-to-text accuracy. Automatic language detection covers widely used languages including English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian, giving the model practical reach across multilingual deployments.

Practical use cases include transcription of long recordings, multilingual content analysis, and voice-driven workflows that can trigger function calls from spoken intents, which is useful for assistants and automation pipelines. A 32k token context window supports audio inputs up to thirty minutes for transcription or forty minutes for understanding, making it feasible to handle meetings, podcasts, or extended interviews in a single pass. Performance on speech recognition has been measured against the FLEURS, Mozilla Common Voice, and Multilingual LibriSpeech benchmarks using average word error rate, giving teams a familiar reference point for comparing transcription quality. Because the weights are openly available, organizations can self-host and adapt the model to specialized domains such as legal, medical, or customer support audio where domain vocabulary matters.