Realtime-Venus vs Qwen-Audio-3.0-TTS: One Reads Your Script, the Other Has to Hear You

The tempting comparison between Realtime-Venus and Qwen-Audio-3.0-TTS is price per hour of audio, and it produces a clean answer that is entirely wrong. Qwen-Audio-3.0-TTS-Plus renders speech at roughly $27.6 per million characters, which works out near $1.66 for an hour of finished audio. Realtime-Venus, the 9B full-duplex system Ant Group's Venus Team built with Tsinghua University, has no price at all because it has no hosted service — but self-hosted it would cost you a GPU node running continuously, which is at least an order of magnitude more per hour. The renderer wins on the arithmetic. The arithmetic is measuring two different products: Qwen-Audio-3.0-TTS takes text and returns audio, and Realtime-Venus takes audio and video and returns a conversation. The question that actually separates them is not what they output. It is whether anything has to talk back.

The input is the whole difference

Qwen-Audio-3.0-TTS, released by Alibaba Cloud's Tongyi Lab on 20 July 2026, is a text-to-speech model exposed as a hosted API with no open weights. Text goes in through an HTTP endpoint and audio comes out. There is no audio input path, no turn to take, no transcript to return, and no concept of being interrupted — the model has never heard anything. Its entire cost model is a printing press: characters in, audio out.

Realtime-Venus, by contrast, is permanently listening. The audio-visual checkpoint carries a SigLIP2 visual encoder for streaming frames and a Whisper-Medium audio encoder feeding a Qwen3-8B backbone, and both checkpoints run a one-second interaction loop that continuously updates conversational state and decides whether to speak or stay quiet. It produces speech through discrete S3 tokens and a streaming flow-matching decoder rather than a separate synthesis stage, and it can return text alongside the waveform.

That is not a feature difference. It is the difference between a component and a system. Put the two side by side and one of them can be dropped into a pipeline that already has a speech recogniser and a language model in front of it. The other one replaces all three.

A worked case makes it concrete

Take a support line. A customer calls, describes a problem badly, pauses mid-sentence to find an account number, gets interrupted by a background announcement, and then asks a follow-up before the agent has finished answering.

A system built on Qwen-Audio-3.0-TTS handles this as three vendors and three bills: a recogniser turns the caller's speech into text, a language model decides what to say, and Qwen-Audio-3.0-TTS speaks it. Each hop adds latency, and each boundary is where prosody dies. The critical failure is structural — the TTS leg cannot hear the pause in the middle of a sentence it is already speaking, so the barge-in has to be handled by something in front of it.

Realtime-Venus handles the same call as one model, no external voice-activity detector, no hand-off. Its Full-Duplex-Bench v1.5 figures — 75% response to interruptions and continuation rates of 97% under backchannels, 88% under other-directed speech and 86% under background speech — are exactly the metrics that describe this scenario. They are also self-reported, from the model's own technical report, with no external reproduction. Read them as a design claim, not a measurement.

Voice quality: one has an Elo, the other has nothing

This is the most lopsided part of the comparison, and it favours Alibaba by a wide margin.

• Independent score — Qwen-Audio-3.0-TTS-Plus sits second on Artificial Analysis's Provider Voice Arena with an Elo of 1,259 over 1,544 samples as of 18 September 2026, behind Cartesia Sonic 3.6 at 1,277 and ahead of Inworld Realtime TTS-2 at 1,247 and Speechify Simba 3.2 at 1,241.

• Where it started — the arena's own release column lists this model as a September 2026 entry, which is the tier as currently scored rather than the 20 July 2026 launch date. Whenever it moved, it has held the top two throughout.

• Realtime-Venus — no arena entry, no third-party voice evaluation, nothing. Its only voice-related figure is a self-reported VoiceBench AlpacaEval score of 4.81 that the report describes as matching the best comparison figure.

• What that means — nobody outside the authors has judged whether Realtime-Venus sounds good. That is not a mark against it; it is simply unknown, and unknown is different from bad. But if voice quality is the deliverable, buying an Elo-rated model beats taking an unrated one on faith.

Language, voices, and expressive control

Alibaba's model is built for breadth of delivery. It ships more than a thousand voices, covers sixteen languages including twenty Chinese dialect regions, and exposes 86 inline tags for emotion and delivery — a control surface you write into the text itself, plus natural-language delivery instructions. Output runs up to 48 kHz, up from 24 kHz in the previous generation, and a single synthesis can run to three minutes. The Flash tier targets roughly 300 milliseconds to first packet for real-time playback; the Plus tier is the quality tier and generates at around sixteen characters per second, which is slow enough to matter in a live product and invisible in batch rendering.

Realtime-Venus documents English and Chinese, ships one reference voice with the weights, and gives you a token-to-waveform decoder rather than a voice library. Expressiveness in the Qwen sense — tags, instructions, a thousand presets — has no equivalent here. What it has instead is the ability to change its delivery in response to what it is hearing, which is a different kind of expressive control and much harder to put on a spec sheet.

The cost of each path, honestly

There is a real caveat on the Alibaba number before anyone budgets with it. The widely quoted $27.6 per million characters does not match Alibaba's own pricing page in every snapshot, and the tiers are priced separately. Check the figure at account level rather than trusting a comparison table, including this one.

Realtime-Venus has no list price because there is nothing to buy. The cost is two 9B checkpoints in BF16 — roughly eighteen gigabytes of weights apiece before activation memory — plus a continuous streaming loop, which means a GPU node that bills by the month whether or not anyone calls. At steady volume that is cheaper per conversation than a metered voice API. At intermittent volume it is far worse, and the crossover point is the only financial question that matters.

There is a third path that a lot of teams will land on, and it is worth naming because it changes the shape of the decision. If you keep a cascade, the only leg that needs to be a hosted model is the middle one — the reasoning. OrcaRouter carries neither Qwen-Audio-3.0-TTS nor Realtime-Venus; both voice layers are outside what we serve, and we would rather say that than imply otherwise. The reasoning leg is what we do cover: nearly 200 text models behind one key at provider list price with zero markup, automatic failover across upstreams so a bad afternoon at one provider does not drop a live call, a routing DSL that composes several models into a single call, and model fusion when a decision deserves more than one opinion. In a cascade that is the leg you most want to be able to swap without a contract negotiation, and the one where a vendor price cut lands the same day rather than at renewal.

Cloning cuts both ways

Qwen-Audio-3.0-TTS can clone a voice from a short reference sample, cross-lingually, and is documented as robust to noisy or reverberant input. That is the single most commercially useful capability in the comparison and the largest compliance surface. A product that can reproduce any speaker from ten seconds of audio has a much longer answer to "whose voice is that, and who consented to it" than one that only ever speaks in a vendor's presets.

Realtime-Venus ships a reference voice with the weights. You can clone with it if you have the audio and the training run, but nothing in the release is built around making that easy — which, for a self-hosted deployment inside a compliance boundary, is arguably the safer default.

Which one you are actually buying

Buy Qwen-Audio-3.0-TTS when the audio is the output. Narration, localisation, accessibility, dubbing, any workflow where the text exists before the sound does and quality per rendered hour is the metric. Start on Flash if playback is live, move to Plus when the voice itself is the product, and price the cloning workflow separately from the preset library.

Reach for Realtime-Venus when the input is a person and the system has to behave like a listener. Camera-aware assistance, hands-free interaction, anything where a human is going to interrupt mid-sentence and expect the system to handle it. The trade is stark: you give up an Elo-rated voice, a thousand presets, and any hosted option, and in exchange you get the only one of the two that can hear you.

Most products want both, in different places. What they should not share is a cost model — price per hour of audio is the right metric for exactly one of these two models, and it is not the one holding the conversation.