Tencent's New Hy Realtime Voice Model Talks, Calls Tools on Its Own and Sings on Request in the Yuanbao App

Tencent's Hunyuan team announced Hy Realtime on November 13, a realtime voice model built for live conversation rather than one-shot questions. It is already running in production: the model powers voice calls in the Yuanbao assistant — selectable under the "Chuyu" voice — and handles real conversations in Tencent's customer service. A developer API on Tencent Cloud is promised soon.

The headline upgrade is intent recovery for the way people actually talk. Tencent says Hy Realtime handles vague phrasing, half-finished sentences and mid-sentence self-corrections by reconstructing what the user meant from conversational context, and holds instructions reliably across long multi-turn exchanges. The agentic part is the bigger shift: rather than always routing through a fixed pipeline, the model decides on its own whether a question needs a web search, unifying the search decision, tool invocation and answer generation into one flow. In other words, it is designed to be an agent you talk to, not a voice front-end for a chatbot.

The most eye-catching feature is singing, developed jointly with the Lingyin AI team at Tencent Music. Users can request songs in natural language — by artist, album, genre, language, release year or popularity — and the model carries the request through multi-turn context before performing it. Tencent says it tuned the singing for expressiveness, breath control and rhythm, and positions the capability as evidence that voice assistants can move past question-and-answer into entertainment.

Tencent published results on four public voice benchmarks, all vendor-reported and not independently reproduced. Hy Realtime leads the Wild Speech Bench at 75.35, ahead of Seed2.0-Lite at 74.80, Gemini-3.8-Flash at 73.61, Qwen3.8-Omni-Flash at 71.36 and GPT-Realtime-2.1 at 71.37. On the other three it lands just behind the leaders: 92.20 on Voice Bench versus Gemini-3.8-Flash's 92.26, 98.60 on Big Bench Audio versus Qwen3.8-Omni-Flash's 99.10, and 85.84 on Open Audio Bench versus 86.85. All four scores exceed the previous generation, Hy2-O. Tencent notes Gemini-3.8-Flash was evaluated in its low think mode. On the company's internal Hy-Eval-Audio3.0 test, Hy Realtime scored 39.33, below Qwen3.8-Omni-Flash at 40.92 and well below Gemini-3.8-Flash at 61.01.

For users, the fastest way to try it is the Yuanbao app: update to the latest version, start a voice call, and pick the Chuyu voice in the settings panel. A web demo hosted on Tencent Cloud covers a subset of the capabilities. Developers will have to wait for the Tencent Cloud API, which the company says is coming "soon" — and which marks the first time Tencent pushes its self-developed realtime voice stack beyond its own products.

The roadmap names three priorities: duplex interaction for natural back-and-forth interruptions, long-term personalized memory, and cooperation between foreground and background agents acting on a user's behalf. No dates were given.

Realtime voice has become one of the most contested fronts in AI, with OpenAI, Google and a crowd of startups pushing voice agents that listen, act and increasingly transact. Tencent's bet leans on distribution: Yuanbao puts the model in front of a huge consumer audience immediately, and the customer service deployment is a live stress test against real users rather than demos. The release follows the November 4 launch of Hy4, Tencent's largest open-weight model, and signals the voice stack is becoming a first-class product line rather than a demo.