Why TTS Models Now Look Like LLMs — Samuel Humeau, Mistral
AI Engineer Europe 2026 · 22:26
AI models, enterprise agents, and compute infrastructure
Mistral AI builds foundation models, applications, developer tools, and compute infrastructure for enterprises and governments. Its Vibe agent, which replaced Le Chat in 2026, handles research, document drafting, and coding through reviewable pull requests. Developers can build through APIs in Mistral Studio, while Forge lets enterprises build models grounded in proprietary knowledge. Deployment control is central to its offering: customers can run models in their own data centers or cloud environments, or use Mistral’s regional inference endpoints in Europe and the US.
Founded in 2023, Mistral is led by co-founder and CEO Arthur Mensch, alongside co-founders Guillaume Lample, Chief Science Officer, and Timothee Lacroix, CTO. Its research includes Mixtral 8x7B, a 2023 open-weight model using a sparse mixture-of-experts architecture. At each layer, a router selects two of eight expert groups for each token, activating only part of the model’s parameters. This approach expands model capacity while limiting the computation needed per token.
Sacra estimated Mistral reached $400 million in annual recurring revenue in January 2026. In 2025, Mistral raised a €1.7 billion Series C, led by semiconductor equipment manufacturer ASML, at an €11.7 billion post-money valuation.
AI Engineer Europe 2026 · 22:26
Devendra Chaplot · Devendra Singh Chaplot
AI Engineer World's Fair 2024 · 18:16
Affiliations reflect their AIE appearances, not necessarily current employment.
Start here for an overview of pre-training and preference tuning, then learn about Codestral’s fill-in-the-middle capabilities and editor integrations.
Devendra Chaplot · Devendra Singh ChaplotAI Engineer World's Fair 2024
Learn how speech recognition, an LLM, and streaming speech generation combine into a voice-agent pipeline.
Samuel HumeauAI Engineer Europe 2026
Chaplot explains how open model distribution can complement commercial offerings, with cloud and private deployment options for different settings.
Humeau contrasts audio bitrate with spoken text’s information rate to motivate efficient speech representations. His discussion connects speech generation to streaming text input and perceived latency.
Affiliations reflect each recorded session, not necessarily current employment.