Ming-3.0-tts-0.5B vs Voxtral 4B TTS: One of These Weights Is Not for Sale

Read the licence before the benchmark, because in this pairing it settles the argument first. Ming-3.0-tts-0.5B is MIT — download it, fine-tune it, ship it inside a commercial product, attribute Ant Group and carry on. Voxtral 4B TT​S is CC BY-NC 4.0, inherited from the reference-voice datasets behind it, which means Mistral AI's weights are not licensed for commercial use at all. The commercial path for Voxtral is the API at $0.016 per thousand characters. So the real question is not which model sounds better. It is whether you are buying a voice you operate or renting a voice you call, and only one of these two checkpoints lets you choose. 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.

That framing sounds like it makes the comparison easy, and it does not, because the two models are good at different things and the hosted one comes with something the open one cannot offer: a published latency and throughput curve, run on real serving infrastructure, at three levels of concurrency. Ming-3.0-tts-0.5B was published to Hugging Face on September 24, 2026 and its inference code to GitHub on October 19, 2026 with no announcement from Ant Group for either — the model is 19 days old as an artefact and hours old as something you can actually run. Voxtral 4B TT​S was announced by Mistral in March 2026 and has had seven months to be measured. Below, the licence, the deployment shape, the numbers each side can honestly claim, and the questions neither vendor has answered.

Two licences that are not variations on a theme

• Ming-3.0-tts-0.5B — MIT. Commercial use, modification, redistribution and fine-tuning all permitted, with attribution. A 2.96 GB checkpoint you can put in a container image, run on your own accelerator, and ship to a customer without asking anyone's permission.

• Voxtral 4B TT​S — CC BY-NC 4.0. Non-commercial. Mistral's card is explicit that the model inherits the licence of its reference voices, drawn from sources including EARS, CML-TTS, IndicVoices-R and a natural-audio Arabic set. The weights download fine; the licence is the constraint, not the file.

The distinction people get wrong is that the API is not under the same restriction. Calling Voxtral through Mistral's service is a commercial transaction with the vendor, and $0.016 per thousand characters is the price. Self-hosting it inside a paid product is a different act, and it is the one the licence forbids. If your deployment has to be on-premises — data residency, air-gapped inference, a customer who will not approve a third-party endpoint — then Voxtral 4B TT​S is not an option at any price, while Ming-3.0-tts-0.5B is available unconditionally.

That is a hard boundary, and it is worth separating from the softer one. A non-commercial licence still permits evaluation, research and internal experimentation, so the sensible way to read this table is that Voxtral is a service you may test for free and Ming is an artefact you may own outright.

What half a billion parameters buys, and what it does not

• Size and architecture — Ming-3.0-tts-0.5B is a dense 0.5B model on a Qwen2.5-0.5B backbone with a diffusion-transformer head, and the checkpoint bundles its audio tokenizer and speaker encoder. Voxtral 4B TT​S is roughly 4 billion parameters, served through vLLM-Omni, built on Mistral's Ministral-3-3B base.

• Output — Ming at 44.1 kHz; Voxtral at 24 kHz in WAV, PCM, FLAC, MP3, AAC or Opus. A 44.1 kHz master is the difference between a file you can hand to a video editor and a file that has to be resampled first, and it matters more in dubbing and broadcast than in a phone agent.

• Modes — Ming has named tasks (TT​S, FREEFORM, VOICE_DESIGN, MULTILINGUAL, BASIC, EMOTION, DIALECT, IP) with documented mutual exclusions. Voxtral does streaming and batch, with no equivalent task taxonomy published.

• Voices — Ming ships more than 100 built-in Chinese voices and five built-in English ones, plus zero-shot cloning from a short clip and voice design from a natural-language description. Voxtral ships 20 preset voices and adapts to a reference clip as short as three seconds. Twenty is a curated roster; a hundred is a cast list. Which you want depends on whether you are building a product with one voice or a tool that lets other people pick one.

• Paralinguistics — Ming's 14 inline tags cover breath, cough, sigh, laughter, hesitation, surprise and questions, plus explicit emotion, speed, pitch and volume controls. Mistral publishes no tag vocabulary for Voxtral. Absence of documentation is not absence of capability, and it is also not evidence of it.

The 0.5B parameter count is the part that gets misread. Ming is not a smaller version of Voxtral; it is a different design, betting that a compact backbone plus a bundled speaker encoder and a tokenizer is enough for controllable speech, and spending its capacity on control rather than on scale. In Ant's own multilingual evaluation it beats VoxCPM2 on content accuracy in four of five languages while losing speaker similarity in all five — which is what a model tuned for words over timbre looks like on a scorecard.

The one number Voxtral has that Ming does not

Mistral published a serving benchmark, and it is the most useful thing in either release for anyone planning capacity. On a single H200 running vLLM-Omni 0.18.0, with a 500-character input and a ten-second audio reference:

• Concurrency 1 — 70 milliseconds of latency, real-time factor 0.103, about 119 characters per second of GPU time.

• Concurrency 16 — 331 milliseconds, RTF 0.237, about 879 characters per second.

• Concurrency 32 — 552 milliseconds, RTF 0.237 carried forward, about 1,431 characters per second.

That is a capacity plan in three lines: latency stays under a second well past the concurrency most voice products actually reach, and throughput scales roughly twelvefold from one stream to thirty-two. It is Mistral's own harness, unreproduced, and it assumes the reference environment — but it is a shape you can size a deployment against.

Ming's equivalent number is independent rather than vendor-reported, and it measures something slightly different. Artificial Analysis records Ming-3.0-tts at 42.1 characters per second of generation time on its own harness, one request at a time. That is roughly a third of Voxtral's single-stream rate by a different method on different hardware, and the honest statement is that the two figures are not directly comparable: one is Mistral's throughput at a chosen concurrency on an H200, the other is a third party's single-stream measurement on unpublished hardware. What can be said is that both are fast enough for interactive speech and neither is a real-time constraint.

The throughput difference that does matter is licensing-coupled. Voxtral's curve is a reason to call the API rather than build around the weights — the serving engineering is already done, and you pay per character instead of per GPU-hour. Ming's rate is a reason to work out your own break-even, because at 42.1 characters per second a million characters is about 6.6 hours of generation, and whether that is cheap depends entirely on what your accelerator costs when it is otherwise idle.

Nine languages against seven, and neither list is a quality claim

• Ming-3.0-tts-0.5B — Chinese, English, German, Spanish, French, Korean and Japanese, with Sichuanese and Cantonese as separate dialect tasks and an explicit language identifier required for the five non-Chinese languages.

• Voxtral 4B TT​S — English, French, Spanish, German, Italian, Portuguese, Dutch, Hindi and Arabic. The card describes "diverse dialects" within that support without enumerating them.

The overlap is five languages: English, French, Spanish, German and one of Korean or Japanese depending on how you count regional coverage. Each model covers something the other does not — Italian, Portuguese, Dutch, Hindi and Arabic on Voxtral's side; Korean, Japanese, Sichuanese and Cantonese on Ming's. If your roadmap includes Arabic or Hindi, the decision is over before the first benchmark is read.

What neither vendor publishes is per-language quality. Mistral gives no error rate for any of its nine languages. Ant tested five non-Chinese languages on the Mi​niMax set and published numbers, but they are Ant's own harness against Ant's own chosen competitors, and they say nothing about Korean or Japanese against Voxtral because Voxtral was not in the table. So the language comparison here is coverage, not competence, and any article that presents it as competence is guessing.

The board that has both, sort of

Both models appear on Artificial Analysis's speech leaderboards, but not on the same capture date and not with the same surrounding field, so treat this as orientation rather than a head-to-head result.

• Ming-3.0-tts-0.5B — Provider Voice 1,188 Elo (interval 1,173–1,203, rank range 17–23, nineteenth row of the list); Controlled Voice 1,069 (1,053–1,085, rank range 10–18, thirteenth row). Both captures are from October 19, 2026, the day its inference code went up.

• Voxtral 4B TT​S — Controlled Voice 1,005 Elo, rank range 24–30, twenty-sixth row of the same October 19 capture. Its provider-voice position was not in the material we could verify.

The Controlled Voice board is the one worth trusting, because every model there clones the same fixed reference voices and the voice cast stops being a confound. On the capture above, Ming's 1,069 and Voxtral's 1,005 sit 64 points apart, and with overlapping rank ranges the honest conclusion is that Ming scores higher on a measurement taken later against a larger field. Do not read it as a settled 64-point win. Do read it as evidence that a 0.5B open checkpoint is not out of its depth next to a 4B vendor model on voice fidelity, which is the surprising part.

What a self-hosted Ming pipeline actually costs you in engineering

Adopting Ming is not a download and a function call. The reference environment is a pinned one: PyTorch 2.6 with CUDA 12.4, tested on an H800-80GB and an H20-96G, flash-attention optional with an SDPA fallback. Text normalisation needs pynini 2.1.6 — a library with a reputation for being awkward to install — or you pass a flag that turns normalisation off and accept that numbers, currencies and abbreviations go to the model raw.

Those are the costs that never appear in a comparison table and always appear in a sprint. Voxtral's equivalent costs are hidden inside the API price: someone else pinned the environment, someone else handles the concurrency, and the trade is that you cannot change the voice cast beyond adapting from a reference clip, and you cannot put the model on a machine you control for commercial purposes.

Ming's maturity also has to be priced honestly. The GitHub repository went up on October 19, 2026, with the first commit carrying the inference code, a Dockerfile, one-click examples and a task cookbook, and documentation commits landing through that afternoon — including one that marked vLLM-Omni support. vLLM-Omni is the same serving stack Mistral used for its Voxtral benchmark, which is a useful coincidence: the path to a production deployment for Ming is the path Voxtral already documented, one layer down.

Where routing fits, and the part it does not touch

OrcaRouter does not serve Ming-3.0-tts-0.5B; it is a checkpoint you run. Voxtral 4B TT​S is available through Mistral's own API, and this piece does not claim we route it. What a router does reach in either architecture is the text around the voice — the script, the SSML-like tag decisions, the translation step that turns one master script into nine languages before synthesis, the summariser that shortens a passage to fit a time slot. Those calls happen hundreds of times per audio minute and they are where a pipeline's cost and its failure modes actually concentrate.

That is the layer we operate: more than 200 text models behind one key at provider list price with 0% markup added, so a vendor's price cut reaches you the day it lands rather than at the next repricing; automatic failover across providers, so a single provider's bad afternoon does not stop the script generation that feeds the voice; and a routing DSL for composing several models into one call when the translation step and the style rewrite need different models. The voice model at the end of the chain is yours to choose, and this article is about choosing it.

What neither card tells you

Three questions decide most real deployments, and neither vendor answers any of them.

• How does quality hold under concurrency? Mistral measures latency and throughput at 32 streams but publishes no quality delta between batch and streaming. Ming's release has no concurrency story at all — a single-request harness, a reference GPU, and no vLLM throughput curve, only a support note.

• What does a failed generation sound like? Neither card documents truncation behaviour, repetition loops, or what happens when a 14-tag script collides with a sentence the model cannot parse. Ant's card is unusually candid about its weak spots — unstable multi-event paralinguistics, emotion control that degrades when the words carry no emotion, French similarity as its worst language, and open-ended voice design that does not match the strongest system it was evaluated against — but a limitations section is not a failure taxonomy.

• Who is maintaining this in a year? Mistral is a company with a commercial TT​S service and an obvious incentive to keep it current; the March checkpoint has already been stable for seven months. Ming-3.0-tts-0.5B has one like, six downloads and a repository created on the day this was written. Ant Group may well support it — the Ming family includes ASR, audio and tiny variants, which suggests a programme rather than a one-off — but nothing in the release commits to anything.

That last point is the one to weigh against the MIT licence. A permissive licence guarantees you the right to use the weights forever; it does not guarantee that anyone will fix a bug in the inference code, and the difference only becomes visible in month seven.

The decision, in the order it actually arrives

Start with the deployment constraint, because it eliminates one option outright. If the model has to run on infrastructure you control and the output has to be commercially usable, the answer is Ming-3.0-tts-0.5B, and the remaining questions are engineering: pin the environment, get pynini installed or accept unnormalised text, budget about 6.6 GPU-hours per million characters, and accept that the voice will be a little less faithful to a reference clip than the best open alternative. If the model has to speak Arabic, Hindi, Italian, Portuguese or Dutch, the answer is Voxtral 4B TT​S through the API, and the remaining questions are commercial: $0.016 per thousand characters, twenty voices, a 24 kHz output that may need resampling, and a vendor who has published the serving curve you will be sizing against.

If neither constraint binds, the tie-break is what your product is. A dubbing or broadcast pipeline that needs many voices, dialect coverage and inline emotion wants Ming's cast and its 44.1 kHz output, and will pay for it in operations. A voice agent or an IVR that needs predictable latency, streaming and someone else on call wants Voxtral's API, and will pay for it per character and in the loss of a voice cast it cannot extend beyond twenty.

The one thing worth doing before either: write down which languages you will actually ship in the next twelve months, and which of them your legal team will allow to leave your infrastructure. Those two lists narrow this from a model comparison to a single candidate faster than any Elo number, and unlike the leaderboards, they are current.