腾讯混元发布实时语音模型 Hy Realtime：自主判断联网搜索、联合 TME 支持点歌演唱，已上线元宝语音通话，腾讯云 API 近期开放

【Key points】

· Launch: On Nov 13 (16:32 Beijing time), Tencent Hunyuan announced Hy Realtime, a new real-time voice model, on its official WeChat account.

· Understanding: it recovers intent from mumbled, half-finished or self-corrected speech, and holds instructions better over long multi-turn talks.

· Tools: new agentic tool calling. The model decides when to search the web, with the search decision, tool call and answer unified in one model.

· Voice: Tencent says quality, pronunciation, prosody and emotion all beat the previous generation.

· Singing: co-designed with TME's Lingyin AI team. Users can request songs by singer, album, genre, language, year or popularity, and the model sings them.

· Availability: live in Yuanbao voice calls (voice "Chuyu") and Tencent customer service, plus a web demo. A Tencent Cloud API is "coming soon", with no pricing yet.

· Evals: internal expert blind tests and unnamed public benchmarks only. No scores were published.

【Our read】

Background

this follows Hy4 (e/1085). Full duplex and long-term memory are listed as next steps.

Comparison

• tool calling is now table stakes.
• Gemini 3.8 Live, Qwen3.8-Omni-Flash-Realtime and Doubao's Seeduplex 3.0 all offer function calling plus web search via public APIs.
• Doubao's O2.0 voice model already sings from a licensed catalog.
• OpenAI (GPT-Live-1.1), SpaceXAI (Grok Voice 3.0, $0.08/min) and ElevenLabs focus on task success, latency and price, not singing.

Impact

consumers can try it in Yuanbao today. Developers must wait for the API, its pricing and independent tests.

【Verifiable】

• · Official demo
• · Official WeChat post (Tencent Hunyuan, Nov 13): "实时语音模型 Hy Realtime 发布"
• · IT Home
• · Yicai
• · NBD