Tencent Gander: open voice AI model with a two-brain architecture

Tencent unveiled Gander, an open model built on a Cerebellum-Brain design: the "cerebellum" handles real-time voice dialogue and interruptions while the "brain" runs background tasks. Weights are on Hugging Face under Apache 2.0, based on MiniCPM-o 4.5 with 9.4B parameters. Task accuracy is about 40% versus 60% for GPT-Realtime, and the full setup needs three GPUs.

• Cerebellum-Brain architecture runs dialogue and background tasks in parallel
• Weights released on Hugging Face under Apache 2.0
• Based on MiniCPM-o 4.5, about 9.4B parameters, up to 19 GB VRAM
• Task accuracy around 40% vs about 60% for GPT-Realtime