Open‑weight TTS wave
Two open‑weight text‑to‑speech releases — Mistral’s Voxtral and Smallest.ai’s Lightning V3 — claim voice quality that challenges commercial offerings like ElevenLabs and are aimed at real‑time, edge, and privacy‑sensitive use cases. Open weights lower vendor lock‑in and make production voice easier to run on mobile or on‑prem. (venturebeat.com, theprint.in)
Mistral published Voxtral TTS on March 26, 2026 and released the model weights and paper under an Apache 2.0 open‑weights license on its site and Hugging Face. (mistral.ai) Voxtral is available in edge‑scale and server‑scale checkpoints (a Mini/4B realtime variant and a Small/24B production variant), with the Small checkpoint needing ~55 GB GPU RAM while the Mini is designed to run in a few gigabytes of RAM for on‑device use. (modelscope.cn) Mistral’s human preference and automatic evaluations report Voxtral surpassing ElevenLabs on speaker‑similarity and naturalness in several benchmarks, achieving zero‑shot voice cloning from ~3–5‑second references and time‑to‑first‑audio comparable to ElevenLabs Flash v2.5 per the company paper and press coverage. (mistral.ai) Smallest.ai announced Lightning V3 on March 27, 2026 as a real‑time TTS tuned for conversational prosody, reporting a 3.89 MOS in its conversational evaluations and per‑metric scores for intonation (3.33) and prosody (3.07), and offering voice cloning from seconds of audio. (smallest.ai) Lightning V3 is distributed via Smallest.ai’s API/SDK and demo spaces (pay‑as‑you‑go with a free starter tier) rather than as a published downloadable weight; Hugging Face demo spaces show ~175 ms latency, 44.1–44 kHz sample rates and ~43 voice presets in published demos. (smallest.ai) The two releases split the product tradeoffs: Voxtral’s Apache‑2.0 open weights and sub‑3‑GB edge variant enable self‑hosting and customization without per‑character API fees, while Lightning V3’s SDK/API and pay‑as‑you‑go model target quick integration, streaming synthesis, and managed scaling for enterprises. (mistral.ai)
