Voxtral 3 TTS vs Voxtral 4B TTS: The Newer Model You Cannot Download

The awkward fact about Voxtral 3 T​TS is that it is better than Voxtral 4B T​TS in almost every way that can be measured and worse in the one way that decides some deployments: you cannot have it. Mistral released Voxtral 4B T​TS on March 23, 2026 as open weights — mistralai/Voxtral-4B-T​TS-2603 on Hugging Face, CC BY-NC 4.0, downloadable and runnable on your own accelerators. Voxtral 3 T​TS arrived on October 20, 2026 as an API product only. The launch page describes EU, US and global endpoints, Studio playgrounds and self-deployment on private infrastructure, and never once says the weights are available. So the seven-month successor is 127 Elo better on the blind-listening board (1,217 against 1,090), 19% cheaper at $13 against $16 per million characters, and about 42% faster at 112 against 79 characters per second — and it is still not a drop-in replacement for the model it supersedes, because the older one is the only member of this family a company can actually own. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Two models, seven months, one licence boundary

That boundary is worth stating precisely before any benchmark, because it reorders the whole comparison.

• Voxtral 4B T​TS — open weights under CC BY-NC 4.0. Non-commercial. You may download it, inspect it, run it, and fine-tune it; you may not ship it inside a paid product. The restriction is inherited from the reference-voice datasets behind it, not from any technical limitation of the checkpoint. Self-hosting it for internal evaluation is fine; self-hosting it in a customer-facing commercial service is not.

• Voxtral 3 T​TS — closed. No weights published as of October 21, 2026. "Self-deployment on private infrastructure or a private cloud" on Mistral's page is an enterprise deployment offer negotiated with the vendor, not a download link, and conflating the two is the easiest mistake to make when reading that page quickly.

The practical consequence: if your reason for wanting Voxtral 4B T​TS was that the audio never leaves your network, Voxtral 3 T​TS does not replace it unless you sign something with Mistral. If your reason was that it was cheaper than a hosted API, Voxtral 3 T​TS replaces it on price and beats it on quality, and the licence question disappears because you are buying a service rather than a file. The same seven-month gap resolves differently depending on which of those two sentences describes your situation.

The scoreboard

• Provider Voice Elo — Voxtral 3 T​TS: 1,217 (#13 of 98), 95% interval 1,202–1,232. Voxtral 4B T​TS: 1,090 (#46), interval 1,078–1,102. Both per Artificial Analysis, read October 21, 2026.

• Price per 1M characters — $13.00 versus $16.00. At roughly 900 characters per spoken minute, $0.70 against $0.86 an hour.

• Throughput — 112 characters per second median versus 79, measured by Artificial Analysis on each model's native host.

• Languages — Voxtral 3 T​TS: 35+ languages and 50+ dialects, presets in 18. Voxtral 4B T​TS: 9 languages — English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic, with dialect support.

• Voice adaptation — Voxtral 3 T​TS: "Pro Voice", 30 minutes of audio. Voxtral 4B T​TS: a 5–25 second reference prompt, described in the model card as little as 3 seconds.

• Weights — closed versus open (CC BY-NC 4.0).

The language line is the biggest change nobody has written about. Voxtral 4B T​TS covers nine languages; the new model claims 35+ and 50+ dialects. That is a four-fold expansion of the product, and it is a stronger argument for the upgrade than the Elo is — 127 points moves you from the middle of the board to the middle of the board, but going from nine languages to thirty-five changes which markets you can sell into.

The voice-adaptation line runs the other way and is the one genuine regression. Voxtral 4B T​TS clones from a short reference clip — as little as three seconds per the model card, with the architecture describing a 5–25 second prompt window. Voxtral 3 T​TS's Pro Voice wants thirty minutes of audio to reproduce a speaker's character, emotional range and style across languages. Thirty minutes is a studio session. Three seconds is a feature. If your product lets a user hear their own voice, the newer model is worse at the job.

What the 4B actually is, under the hood

The parameter count in the old model's name describes a real architecture, and it explains both its strengths and why the successor is not simply a bigger version of it.

Voxtral 4B T​TS is roughly four billion parameters in three parts: a 3.4B transformer decoder backbone built on Ministral 3B, a 390M flow-matching acoustic transformer, and a 300M neural audio codec. It is autoregressive over a semantic VQ codec with an 8,192-token vocabulary, feeding a flow-matching head that predicts acoustic FSQ features across 36 dimensions at 21 levels, running at 12.5 Hz. In plain terms: a small language model reads the text and plans, a flow-matching network turns that plan into audio features, and a codec turns features into samples. It generates up to two minutes natively, with longer output handled by interleaving in the API.

Mistral's published figure for the March model was 70 milliseconds of model latency for a 10-second voice sample and 500 characters, at a real-time factor of about 9.7x. Those numbers were measured on Mistral's serving stack, not independently, and they describe a model that was genuinely fast for its generation. The October model's throughput on Artificial Analysis, 112 characters per second against 79, suggests the successor is faster still — but Mistral publishes no time-to-first-audio for Voxtral 3 T​TS at all, which is a strange omission for a company whose previous T​TS model led with a latency figure.

Two generations of architecture are also visible in the language count. Nine languages in March, thirty-five-plus in October, with the family's transcription models claiming all 24 official EU languages. Whatever changed between the two checkpoints changed the training data and the tokenizer's coverage far more than it changed the acoustic stack.

Who should keep the 4B, and who should move

Keep Voxtral 4B T​TS if any of these is true. You need the audio to be processed on hardware you control and you cannot negotiate a private deployment. You are doing research or evaluation, where CC BY-NC 4.0 permits use and the licence costs nothing. You need short-reference voice cloning — three seconds against thirty minutes is not a close call. You are working in one of the nine languages it covers and the quality is sufficient, in which case a free checkpoint you already run is worth more than a better one you have to call.

Move to Voxtral 3 T​TS if any of these is true. You are selling the output and were therefore never allowed to self-host the 4B commercially anyway — in which case you were already on the $0.016 per thousand characters API, and the new $13 per million is 19% cheaper for a 127-Elo better model, which is the easiest upgrade in this piece. You need a language outside the original nine. You need the agent stack: Voxtral 3 T​TS ships with a LiveKit-based voice-agent SDK, SIP and WebRTC support, sub-second agent turns and a per-minute bundle price of $0.04 with Mistral Small 4 reasoning, none of which existed around the March model.

There is a third group that should run both, and it is larger than it looks: teams that prototype on the open checkpoint and serve on the API. Voxtral 4B T​TS gives you a local model to test prompts, voices and failure modes against with no per-call cost and no data leaving the building; Voxtral 3 T​TS gives you the production endpoint with the language coverage and the licence to sell. The two share a family lineage but not an API surface, so plan for the prompts to need re-tuning — a 127-point quality gap is a different model, not a version bump, and direction controls that behaved one way on the 4B will not necessarily transfer.

If the production path needs to survive a vendor outage, that is the argument for putting the synthesis call behind a router rather than calling Mistral directly. OrcaRouter gives one OpenAI-compatible endpoint across 200+ models with provider list price passed through at 0% markup, so an outage on one synthesis backend fails over instead of going silent — and a voice pipeline that renders overnight batches is exactly the workload where a dead endpoint costs a night rather than a request. Mistral's models are reachable through Mistral's own API and several third-party platforms; check the catalogue for current availability before designing around a specific model ID.

The question Mistral has not answered

Will Voxtral 3 T​TS get open weights? The March model did, under a non-commercial licence, and it became the reference point for what an open T​TS checkpoint from a European lab looks like. Nothing in the October launch repeats that. The page is explicit about API access and private deployment and silent about downloads, which is the same pattern Mistral used for several of its text models before releasing weights months later.

If weights do land, the comparison changes completely: the licence terms will decide whether the open model is commercially usable, and the precedent — CC BY-NC 4.0 — suggests they will not be. That would leave the situation roughly as it is now, with the downloadable model being a research artefact and the usable one being a service.

The other unanswered question is whether the 4B will keep its API listing. Mistral's speech pricing table lists the new family and does not mention the March model, while Artificial Analysis still tracks it at $16 with an active host. Historically Mistral has kept older speech models callable for a long tail; if that holds, teams have a genuinely useful fallback — a slightly worse, slightly more expensive model that is nonetheless a different deployment. If it does not, anyone who built on voxtral-mini-tts-latest in March needs to check which model that alias now resolves to before their next invoice.

The short version: Voxtral 3 T​TS is the better model and the worse asset. Buy it for languages, quality and the agent stack; keep Voxtral 4B T​TS for local work, short-reference cloning and the nine languages it already does well. The one migration that is unambiguously correct is from the 4B's hosted API to the new one — same vendor, same licence posture, cheaper and better — and it is the only move in this comparison that costs nothing to make.