Mistral has released Voxtral 3 with a development kit linking speech recognition, reasoning, tool calls and spoken replies. Its streaming design lets stages overlap and supports interruptions during a conversation. Agents can reach external systems for tasks such as checking invoices or booking appointments, while deployment options include the company’s hosted service.
Mistral, the AI model developer, released Voxtral 3 on 20 October with a software development kit for voice agents. The kit combines transcription, reasoning, tool calls and speech generation in a streaming pipeline. Stages can work before the preceding stage finishes. During a live conversation, an agent can reach external systems to check an invoice or book an appointment.[1], [2]
Voxtral Small Transcribe 3 distinguishes speakers and can use supplied names and technical terms as context. Mistral says the transcription family supports more than 50 languages. The realtime component uses acoustic and semantic information to estimate when a person has finished speaking. The pipeline also supports stopping the agent’s speech when the caller interrupts.[1]
The kit is built on LiveKit, software used for realtime communication. Listening, reasoning and speech output remain separate model stages, allowing a developer to change a model, voice or language. Web search and retrieval from an organization’s systems can run during a call. Mistral offers hosted access; organizations seeking the complete stack on their premises or in a private cloud are asked to contact the company.[1]
