Introduction

Tencent Hunyuan launched Hy Realtime, a new real-time voice conversation model, on November 13. The announcement went out on the official Tencent Hunyuan WeChat account at about 16:32 Malaysia time (08:32 UTC).

Hy Realtime is a closed model. You can use it today in two places: voice calls in Tencent's Yuanbao app (pick the "Chuyu" voice) and a web demo on Tencent Cloud. It also already runs Tencent's customer-service lines. Tencent says the model is "coming soon" to Tencent Cloud as an API for developers.

This release isn't about raw speech quality. Hy Realtime adds two things to a live voice call: tool calling (it decides on its own when to search the web) and singing on request. Tencent built the singing with Tencent Music's TME Lingyin AI team.

What Shipped

  • Better at messy speech: Tencent says the model works out what you meant from mumbled, half-finished or mid-sentence-corrected speech by using the conversation context. It says multi-turn instruction following and context tracking also hold up better in long calls.
  • Agentic tool calling: during a call, the model decides from context whether a question needs a web search. Tencent describes the search decision, the tool call and the spoken answer as one unified step, not a separate search pipeline attached to the model.
  • Singing in the conversation: you can ask for a song by artist, album, genre, language, release year or popularity, combine those filters, and refine the pick in follow-up turns. Tencent's example: "Sing an English folk song from between 2010 and 2020." The model works out the request, matches a song and sings it.
  • Voice output upgrade: Tencent says Hy Realtime beats its previous voice model on sound quality, pronunciation, prosody and emotion. It also says emotional shifts within a conversation sound more natural.
  • Live surfaces: Yuanbao voice calls ("Chuyu" voice, latest app version), Tencent customer service, and the Tencent Cloud TRTC web demo. Tencent says the demo covers "part of" the model's abilities.
  • Access: closed weights. There's no Hugging Face repo, paper or model card for Hy Realtime itself, and no API pricing yet.

Where to Run It Today

  • Yuanbao app: update to the latest version, start a voice call and pick the "Chuyu" voice. Try correcting yourself mid-sentence ("Saturday… no, make that Sunday"), ask something that needs fresh information, or ask for a song by year and genre.
  • Web demo: rtcube.cloud.tencent.com/component/hunyuan is Tencent Cloud's TRTC "Hunyuan voice conversation" page. It went live on launch day.
  • Developers: wait for the Tencent Cloud API. Tencent hasn't given a date, model ID or price.

A song request worth trying once you're in a call:

Sing me a mellow English folk song released between 2010 and 2020.
Something a bit more upbeat from the same years, please.

Numbers

Tencent shared few hard numbers with the launch:

  • Text-to-speech data retention: Tencent rebuilt how it produces training data so more of its text model's knowledge and reasoning carries over into the voice model. The share of high-quality text data successfully turned into speech-task data went from 21% to 69%, about 3.3×. That figure is from Tencent's launch post as reported in launch-day coverage; we couldn't load the WeChat post itself to check it.
  • Evaluations: Tencent cites internal expert A/B tests for singing, internal blind tests for search-and-answer quality in Chinese and mixed Chinese-English use, and an internal "IQ" test set covering maths, science, reasoning, knowledge, complex instructions and multi-turn chat (non-thinking mode). It also says it leads "several public benchmarks" for voice interaction. None of these came with benchmark names or scores we could verify.
  • No latency figure was published.

Caveats

  • Not full duplex yet. Tencent lists duplex interaction, long-term personal memory, and foreground/background agent handoff as next steps. That puts Hy Realtime behind open full-duplex models like Haru-4B and inclusionAI Duo v1 on talking and listening at the same time.
  • China-first surfaces. Yuanbao and the TRTC demo are built for Chinese users, and Tencent's own evaluations focus on Chinese and mixed Chinese-English speech. English singing is in the examples, but there's no data on other languages.
  • No API, no price. The Tencent Cloud API is announced but not live, so you can't build on it today.
  • Closed and undocumented. There's no technical report, parameter count or architecture detail for Hy Realtime. Don't confuse it with Tencent's Hy ASR 3.0 report or the open Hy-SimulMT translation model. Those are separate releases.
  • Song licensing is unclear. Tencent describes matching a requested song and then singing it. It hasn't said what catalog that draws on or how rights work outside its own apps.

Conclusion

Hy Realtime is Tencent's case that a voice assistant should act and perform, not just answer. It runs web search from inside the call and takes song requests by artist, genre or year. It also comes with a reported jump in how much of the text model's ability survives in the voice model. It's closed, there's no API to build on yet, and true full-duplex is still on the roadmap. Still, singing on demand inside a live call is something most real-time voice stacks can't do today. Try it in Yuanbao with the Chuyu voice, and keep an eye on Tencent Cloud for the API.

—Anabel ♡