Tencent unveils Gander model that talks while handling background tasks
Tencent's Hunyuan Speech team introduced Gander, a model that can keep a real‑time conversation going while a separate "brain" component works on complex tasks. The architecture splits responsibilities: a fast "cerebellum" manages speech, image and text input second‑by‑second, and a swappable "brain" handles reasoning, code generation or other agent functions. Users may interrupt at any moment,…
Key points
- Gander interrupts users in only 8% of 100 Full‑Duplex‑Bench v3 scenarios, versus 13.5% for GPT‑Realtime.
- Model processes speech, images and text simultaneously using a “cerebellum” for conversation and a swappable “brain” for complex tasks.
- Tencent plans to open‑source Gander’s weights and training data; a GitHub repository already hosts the code.
In the Full‑Duplex‑Bench v3 evaluation covering 100 scenarios, Gander interrupted users in only 8 % of cases, compared with 13.5 % for OpenAI's GPT‑Realtime and about 48 % for the weakest competitor. The model lagged behind on overall task accuracy, especially in video and audio understanding, which the researchers attribute to training that favoured fluid conversation over precise perception. Gander was trained on roughly 2.7 million examples. Tencent plans to publish the model weights and training data after completing its open‑source release process, and a GitHub repository already hosts the code and demos.
Model pages: Gemini Live → · Claude Code →
Tencent's Gander aims to keep talking while it works in the background
The Decoder · 20 September 2026
Loading the full article…
This text was published by The Decoder and written by Jonathan Kemper. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Tencent releases Hy4 open-weight model, claims top-tier Chinese AI · 2 src
- Seven open specialist models under 10B parameters outperform larger systems in niche tasks · 1 src
- oMLX 0.7.4 adds prefill accelerator using a second Mac for large models · 1 src
- OpenAI’s 2022: the year ChatGPT and AI art sparked debates over usage and ethics · 1 src
- Axion1-0/Axion-1.0-35B · Hugging Face · 1 src
Comments
via GitHub Discussions