Tencent Hunyuan releases Hy Realtime voice model with singing and web search

Overview

Updated 43 minutes ago

Tencent's Hunyuan team released Hy Realtime on November 13, a voice model that listens and speaks inside the same call. It is already handling Yuanbao voice chats and Tencent customer service calls. Two new abilities stand out: it decides mid-conversation whether to search the web, and it will sing a song on request.

Where it lands matters more than what it does in a demo. Tencent is running the model on its own products first, then plans to sell access through Tencent Cloud. That gives Chinese companies a domestic option for real-time voice agents, a market where OpenAI, Google, and Alibaba already sell competing models.

Why it matters

Tencent is running a real-time voice agent in its own apps, then selling it to Chinese firms that want a domestic alternative to OpenAI.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

69%

High-quality text data retained in speech tasks

Tencent rebuilt its text-to-speech data pipeline and raised the effective retention rate from 21% to 69%.

98.60

Big Bench Audio score

Second to Qwen3.8-Omni-Flash at 99.10 in Tencent's own comparison chart.

75.35

Wild Speech Bench score

First place in Tencent's chart, ahead of Seed2.0-Lite at 74.80.

39.33

Self-built IQ test score

Third behind Gemini 3.8 Flash at 61.01, but double the previous Hy2-O score of 19.54.

2.52

Product expert evaluation

Slightly below Doubao's 2.56 in Tencent's own expert review.

Voices

Curated perspectives — historical figures and your fellow readers.

Organizations Involved

Tencent Hunyuan

AI research and product unit

Released Hy Realtime on November 13, 2026

Tencent's in-house AI model team, which builds the Hunyuan family of language, speech, and multimodal models.

Tencent Music Entertainment

Music streaming and AI research

Co-built Hy Realtime's singing feature through its Lingyin AI team

Tencent's music arm, whose Lingyin AI group worked with Hunyuan on song selection and performance.

Yuanbao

Consumer AI assistant app

First product to carry Hy Realtime voice calls and singing

Tencent's consumer AI assistant, which runs Hy Realtime in its voice-call feature.

Timeline

3 events Latest: Today

November 2026
Tencent releases Hy Realtime voice model
Today Product launch

The model goes live in Yuanbao voice calls and Tencent customer service, adding web-search tool calls and singing.
Today
Web demo opens with 10-minute session cap
Today Product

The demo exposes web search and AI noise reduction, and requires phone verification to enter.
Today
Tencent Cloud API promised 'soon'
Today Statement

Tencent gives no pricing, concurrency limits, or voice options for the developer API.
Today

Scenarios

Tencent Cloud publishes Hy Realtime API pricing and opens self-serve access

Likely Resolves by Q1 2027

Discussed by: Tencent's launch announcement and Cocoloop's coverage both flag the API as coming, with pricing unstated.

Tencent said the developer API will arrive through Tencent Cloud 'soon.' If it follows the pattern of its other model APIs, it will post a product page with per-unit pricing and rate limits. That page is the signal that outside developers can build on the model rather than only use it inside Tencent apps.

Singing stays a Yuanbao feature and never reaches the developer API

Possible Resolves by Q2 2027

Discussed by: Cocoloop notes Tencent did not say whether singing will reach the API or how song copyright will be handled.

Singing depends on licensed music and a partner team at Tencent Music. Those constraints may keep it inside Tencent's own apps, where the company controls rights and distribution. Developers building customer service or hardware products would then get the voice model without the music feature.

A Chinese customer-service or hardware vendor ships a product built on Hy Realtime

Possible Resolves by Q2 2027

Discussed by: Cocoloop and Geeknetic both describe customer service, coaching, in-car, and smart-hardware as the target markets.

Tencent's stated path is to refine the model on its own call volume, then sell it to companies building service and device products. A named outside vendor announcing a shipping product would show the model cleared the pilot stage. Without one, Hy Realtime remains a Tencent-internal tool.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

May 2018

Google Duplex (2018)

Google demonstrated an assistant that called restaurants and salons to book appointments, handling pauses and interruptions in real time. The demo drew attention for sounding human and for the awkwardness of a machine calling businesses without saying so.

Then

Google said the assistant would identify itself, and limited early tests to a small set of businesses.

Now

Duplex remained a narrow feature rather than a general voice agent, showing how hard it is to move from a controlled demo to open conversation.

Why this matters now

Hy Realtime is also a real-time voice agent moving from demo to product. Duplex shows the gap between a polished demonstration and a service that handles unpredictable calls.

May 2024

GPT-4o real-time voice (May 2024)

OpenAI showed a voice mode for GPT-4o that responded with low latency, changed tone, and sang in a demonstration. The company delayed the rollout to address safety and reliability.

Then

Voice mode reached paying users months after the demo, with limits on some capabilities.

Now

Real-time speech became a baseline expectation for frontier AI assistants, and competitors began publishing latency and speech benchmarks.

Why this matters now

Tencent's benchmark chart compares Hy Realtime to Gemini, Qwen, Seed, and GPT realtime models. GPT-4o set the standard those comparisons are measured against.

October 2024

OpenAI Realtime API (October 2024)

OpenAI opened its real-time speech models to developers through an API, letting companies build voice agents without stitching together separate transcription, language, and speech-synthesis services.

Then

Developers began building customer-service and assistant products on the API, with pricing published per unit of audio.

Now

The API model became the standard route for selling voice AI: run it in-house first, then rent it to outside developers.

Why this matters now

Tencent is following the same sequence with Hy Realtime, running it in Yuanbao and customer service before opening a Tencent Cloud API. Whether it publishes pricing will decide how quickly outside developers can plan around it.