Tencent Hunyuan Team Releases Gander, a 9B Apache 2.0 Omni-Interaction Model

Tencent’s Hunyuan Speech Team, with collaborators at Zhejiang University, Shanghai Jiao Tong University, CUHK, and NTU, released Gander as an open 9-billion-parameter omni-interaction model. Weights are published on Hugging Face under Gander-Omni/Gander with an Apache License 2.0 card, derived from OpenBMB’s MiniCPM-o 4.5 base. The accompanying technical report is listed as arXiv:2609.08977, and the serving runtime lives at github.com/Omni-Interaction-Gander/Omni-Interaction-Agent.

Gander is built for continuous spoken and visual interaction while longer tasks keep running asynchronously. The technical report (arXiv:2609.08977) frames a Cerebellum–Brain style split. Separately, the Hugging Face card describes a realtime multimodal Thinker for perception, interruption, and short replies, a streaming Talker for speech synthesis, and the Gander runtime for task lifecycle, follow-ups, and completion.

Streaming contract and components

Interaction is organized into causal one-second units. Per the model card, the Thinker may emit up to eight lexical tokens per speaking unit; the Talker budgets 50 S3 speech tokens per unit and can stream roughly 0.5-second chunks. Audio input is 16 kHz; speech output is 24 kHz; released checkpoints use BF16. The repository layout separates thinker/ and talker/ directories plus a release_manifest.json that records the shared temporal-unit contract.

The runtime README path in the card points operators to clone the Omni-Interaction-Agent repo, point configs at MiniCPM-o 4.5 plus the released Thinker/Talker checkpoints, and run the included serve script—often with the Talker on a second GPU. Dataset release is still marked “coming soon” on the card.

Primary sources are the Hugging Face model card for Gander-Omni/Gander, arXiv:2609.08977, and the Omni-Interaction-Agent GitHub repository.