Latent Space Deep Dive: Mistral Cofounders on Voxtral TTS Architecture, Forge, Leanstral, and Mistral 4 Roadmap

Latent Space·medium signal

Latent Space interviews Mistral cofounders Pavan Kumar Reddy and Guillaume Lample on their full-stack voice strategy. Voxtral TTS (4B params, 70ms latency, 62.8% listener preference over ElevenLabs) completes a pipeline with Voxtral Transcribe (STT) and Mistral's reasoning layer. New details on Forge (enterprise model customization), Leanstral (edge deployment), and the Mistral 4 roadmap — positioning as the only lab offering an open-weight, full speech-to-speech pipeline for enterprise voice agents.