Mistral open-sources Voxtral: a top-tier text-to-speech model you can run yourself

In March 2026 Mistral released Voxtral TTS — a 4-billion-parameter open-weight text-to-speech model published on Hugging Face. It runs on a single 16GB consumer GPU, clones a voice from a few seconds of reference audio, and supports nine languages. Mistral's own blind listening tests preferred it over ElevenLabs Flash v2.5 roughly 63-70% of the time.

The weights are CC BY-NC 4.0 — free for non-commercial use, commercial use via Mistral's API. At ~$0.016 per 1,000 characters, that's a fraction of ElevenLabs' commercial rates. The vendor-run benchmark figure (63-70%) is Mistral's own number — treat it as directional, not certified. Independent reviewers found it genuinely competitive. Months after release, Voxtral remains the standout open-weight TTS: the model that unlocked self-hosting for regulated sectors and set a pricing floor the proprietary market now has to beat.

The regulated-sector angle is the quiet win. Healthcare, finance, and legal deployments often cannot send audio to a third-party API — a compliance requirement, not a preference. Voxtral's self-hostable architecture unlocks that market. No proprietary TTS API can match that, by design. The pricing compression effect — open weights making proprietary APIs cheaper, as seen post-Llama in text generation — is the secondary watch item for the next 12 months.