Voxtral 3 TTS vs Qwen-Audio-3.0-TTS-Plus: Two Cheap Voices, Two Different Maps

Both of these models are cheaper than the frontier and both are outside it, which makes the choice between Voxtral 3 T​TS and Q​wen-Audio-3.0-T​TS-Plus a question about coverage rather than about quality. On Artificial Analysis's Provider Voice board they are 47 Elo apart — Q​wen at 1,264 (#7), Mistral's new model at 1,217 (#13) — and their prices are $19.31 and $13.00 per million characters. Neither gap is decisive. What is decisive is that A​libaba's model was built around Chinese, including twenty dialect regions and eighty-six inline delivery tags, while Mistral's was built around the twenty-four official languages of the European Union. If your product speaks Mandarin, Cantonese or Hokkien, the argument ends at the language list. If it speaks Portuguese, Polish or Dutch, it ends there too, in the other direction. 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.

The numbers, and how much weight they carry

• Provider Voice Elo — Q​wen-Audio-3.0-T​TS-Plus: 1,264. Voxtral 3 T​TS: 1,217. Both per Artificial Analysis, October 21, 2026.

• Controlled Voice Elo — Q​wen-Audio-3.0-T​TS-Plus: 1,125 (#7). Voxtral 3 T​TS: 1,066 (#14). The gap widens from 47 to 59 points once both are cloned onto the same eight reference voices, which means A​libaba's edge is in the synthesis engine, not in its voice library.

• Price per 1M characters — $19.31 for Q​wen-Audio-3.0-T​TS-Plus, $13.00 for Voxtral 3 T​TS. At roughly 900 characters per spoken minute that is $1.04 versus $0.70 an hour.

• Languages — A​libaba: 16 languages plus 20 Chinese dialect regions. Mistral: 35+ languages and 50+ dialects, with presets in 18 languages.

• Voice cloning — A​libaba: reference-audio cloning including cross-lingual timbre transfer. Mistral: "Pro Voice" adaptation from 30 minutes of audio. Neither offers the ten-second instant clone that Cartesia sells.

• Weights — closed on both sides. Both are API-only products.

• Latency — A​libaba publishes a first-character streaming figure for its recognition models and a realtime voice model separately; Mistral publishes no time-to-first-audio for Voxtral 3 T​TS. Neither number is comparable across vendors, and for synthesis neither is an independent measurement.

One caveat on the Elo figures: Voxtral 3 T​TS had been on the board for one day when these were read, with 1,619 comparisons against Q​wen's 2,399. A 47-point gap inside overlapping confidence intervals at that sample size is a direction, not a verdict. Read it as "Q​wen is probably somewhat ahead" rather than as a measured 47.

Where A​libaba's model wins, and it is not a marginal win

Chinese coverage is the first and largest difference, and it is worth being precise about what "20 dialect regions" means. A​libaba's speech line is built for a domestic market where a single written language maps onto speech that a Mandarin-only model mangles: Cantonese, Sichuanese, Wu, Minnan and more. Q​wen-Audio-3.0-T​TS-Plus carries those regions natively, and A​libaba's 3.1 generation added cross-lingual timbre transfer — one voice carried across Mandarin, Cantonese, English and Japanese — which is the feature that matters if you are localising a single presenter rather than casting per market.

The second is delivery control at the token level. Q​wen-Audio-3.0-T​TS-Plus inherited a documented set of 86 inline delivery tags from its generation, letting a script carry emotion, pace and emphasis markers directly in the transcript. Mistral's Voxtral 3 T​TS exposes prompt-level control of emotion, pacing and emphasis plus phoneme-level pronunciation and a vocabulary library — comparable in intent, but A​libaba's tag vocabulary is the more granular of the two on paper.

The third is that A​libaba shipped a cheaper sibling. The 3.1 refresh made T​TS a Flash-tier SKU, so a team that finds the 3.0-Plus tier too expensive has a documented downgrade path inside the same vendor, at roughly the same price point as Voxtral 3 T​TS. Mistral's answer to "cheaper please" is the same model or a different vendor.

Where A​libaba's model is weaker is equally clear: 16 languages is a small map. If your catalogue runs in Polish, Turkish, Vietnamese or Portuguese, Q​wen-Audio-3.0-T​TS-Plus does not cover you and no amount of Elo closes that.

Where Voxtral 3 T​TS wins

Language breadth outside Asia, straightforwardly. Thirty-five-plus languages with 50-plus dialects claimed, and a family pitched at European procurement: Mistral states all 24 official EU languages across the transcription models and EU, US and global API endpoints, with self-deployment on private infrastructure for buyers who need to control where audio is processed and whether it is retained. For a European public-sector or healthcare buyer, that is not a tiebreaker, it is a requirement.

The second is the surrounding stack. Voxtral 3 T​TS is not sold as a lone endpoint: it comes with a LiveKit-based agent SDK supporting SIP, WebRTC and IVR, an agent-turn latency claim under one second, and a per-minute price for the whole loop — $0.04 with Mistral Small 4 reasoning, $0.08 with a different reasoning model. A​libaba has a realtime voice model too, but Mistral's is documented as a bundling proposition with a published per-minute number, which is easier to build a business case on than a per-token rate card.

The third is price, narrowly. $13 against $19.31 is a 33% discount, and at scale that is a real number: ten thousand hours a month is about $7,000 against $10,400. It is not a repricing event, but it is not nothing either.

The routing case is stronger here than in most matchups

Most model-versus-model pieces end with "pick one". This pair is a genuine portfolio problem, because the two coverage maps barely overlap at the edges and both are cheap enough to run side by side. A European e-commerce platform narrating product pages in eight EU languages and also serving a Cantonese-speaking customer base needs both. So does a dubbing house working on a series with Mandarin dialogue and French subtitles.

The awkward part of running both is not the model choice, it is the plumbing: two vendors, two auth schemes, two billing relationships, two sets of failure modes. OrcaRouter exists for exactly that shape — one OpenAI-compatible key across 200+ models, provider list price passed through at 0% markup so a vendor price move is live the same day, and a routing DSL that can send a language-tagged request to the model that covers that language. Neither Voxtral 3 T​TS nor Q​wen-Audio-3.0-T​TS-Plus is in OrcaRouter's catalogue as of this writing, and one API for 200+ models is a claim about the catalogue we do run, not about these two — today both are reached through their vendors' own APIs and several third-party platforms. The design point stands regardless: if your language matrix requires two vendors, you want the split to live in a routing rule rather than in two code paths.

What neither vendor has published

Three gaps, and they are the same three in both cases, which says something about how young this tier of the market still is.

No per-language scores. Both Elo figures are aggregates. A model can sit at 1,264 overall and be excellent in Mandarin and mediocre in French; the board cannot tell you, and neither vendor publishes a breakdown by language. For any decision that turns on quality in a specific language, the only reliable test is your own corpus, read by a native speaker who is not told which model produced which sample.

No long-form endurance figure. Arena clips run seconds. Neither model documents a maximum generation length for its T​TS endpoint, and neither has a published test of coherence across ten minutes — the length at which an audiobook either works or does not. Q​wen's 3.1 generation moved long-form scene generation into a separate -Next model rather than extending the T​TS endpoint, which is an implicit admission that the base model is not the long-form tool.

No independent latency measurement. Mistral publishes nothing for Voxtral 3 T​TS synthesis; A​libaba publishes vendor figures. Artificial Analysis reports throughput of 112 characters per second median for Voxtral 3 T​TS, but not for Q​wen-Audio-3.0-T​TS-Plus on the same page. In a live voice agent, time-to-first-audio decides the product experience more than Elo does, and it is the least-documented number in the category.

The honest split

Choose Q​wen-Audio-3.0-T​TS-Plus if the product speaks Chinese — Mandarin with regional accents, Cantonese, or code-switching between Chinese and English — or if you need granular in-transcript delivery tags and want a vendor with a documented cheaper tier beneath the flagship. It costs 48% more than Voxtral and is measurably ahead on both boards; for Chinese-first work that is money well spent.

Choose Voxtral 3 T​TS if the product speaks European languages, if data residency in the EU is a procurement requirement, or if you are building a voice agent and want synthesis bundled with an agent SDK and a published per-minute price. You give up 47 Elo and gain a language map that covers the markets you sell in.

Choose both — and route between them — if your audience genuinely spans both maps. On an Elo-per-dollar reading the two are close enough that the tie should break on coverage, not on score, and a product with customers in Guangzhou and Gdańsk has no single right answer.

The thing worth watching is whether either vendor publishes a per-language breakdown. The moment one does, this comparison stops being an argument about maps and becomes an argument about measurements, which is a better argument to have.