Voice agents this week: callable stacks from Voxtral and Eleven, versus Dessa’s research preview
Mistral’s Voxtral 3 and ElevenLabs’ Eleven Realtime ship callable voice-agent stacks/APIs; Block’s Dessa stresses a speech-only full-duplex research preview with no public API or weights yet. All three speak to realtime talking agents—at different channel maturity.
Mistral launched the Voxtral 3 series: transcription across 50+ languages, expressive TTS, and a voice-agent SDK that chains listen, reason, tool use and speak in realtime, via an EU-hosted API or self-deploy. That is a developer stack and price channel—not the same as a single closed end-to-end assistant.
ElevenLabs shipped Eleven Realtime v1 Mini and v1 Max as complete speech-to-speech voice-agent models with 90+ languages and a large in-library voice set, citing third-party Artificial Analysis figures for accuracy and time-to-first-audio. It is likewise an integrable product-API path that collapses the stack into whole models—still a commercial voice-platform release.
Block introduced Dessa as a from-scratch, speech-only (no text pipeline) true speech-to-speech full-duplex model that can run locally on a MacBook Pro M4 Max or better. It is a research preview/waitlist: no public API, weights or parameter counts were published, and AI safety researchers are invited for early access. Versus the two callable stacks above, Dessa is still a preview channel—not a generally available voice product.
Sources
- Voxtral 3 — Mistral AI
- Introducing Eleven Realtime: a simpler way to build natural voice agents — ElevenLabs
- Dessa - the Conversation Model trained only on speech — Block