What hardware and setup does Gander need to run?

Gander ships as a local server. You clone the repository, build the conda environment from the included environment.yml, download the Gander-Omni/Gander checkpoints from Hugging Face into checkpoints/Gander, then run ./scripts/serve.sh and open http://127.0.0.1:8000. The stack also pulls faster-whisper-large-v3 for speech recognition alongside the Thinker and Talker weights.

How does Gander score against other speech models?

On SpokenQA, Gander reaches 75.60% on Llama Questions and 59.30% on Web Questions. On Full-Duplex-Bench v3 it records 100% turn-taking accuracy with 8% premature interruptions, and it scores 78.53% on the Daily-Omni multimodal understanding benchmark. The technical report on arXiv carries the full comparison tables.

Can I swap out the reasoning agent behind Gander?

Yes. Gander's back brain is pluggable — Codex is the default provider, but the orchestration runtime is what connects the streaming front end to whatever agent handles long-horizon work. That separation is the point of the Cerebellum-Brain design, and it is why the conversational half stays responsive while the reasoning half takes its time.

Is Gander free to use commercially?

Gander is released under the Apache License 2.0, which permits commercial use, modification and redistribution with attribution. Note that Gander is derived from the MiniCPM-o 4.5 base model and relies on faster-whisper-large-v3 for transcription, so anyone shipping it commercially should check those components' own terms as well.

Tencent Hunyuan Speech Team · 2026-09-09 · major

Gander — an open 9B model that listens, watches and works at once

Gander is an open 9B omni-interaction model that takes streaming video, speech and text together. You can interrupt it mid-sentence, and a separate reasoning agent keeps working on long tasks in the background.

Gander omni interaction agent repository on GitHub — GitHub

Gander pairs a fast streaming speech model with a slower reasoning agent, so a voice conversation keeps flowing while long tasks run.

Quick facts

Maker	Hunyuan Speech Team, Tencent (with ZJU, SJTU, CUHK, NTU)
Size	9B parameters
License	Apache-2.0
Base model	MiniCPM-o 4.5
Weights	Gander-Omni/Gander on Hugging Face
Released	Models, code and data

What is it?

A "Cerebellum-Brain" split is the idea behind Gander, an open 9B omni-interaction model released on 9 September 2026 by Tencent's Hunyuan Speech Team with researchers at Zhejiang University, Shanghai Jiao Tong University, CUHK and NTU. The front half handles real-time audio and video and speaks; the back half handles long-horizon reasoning. Weights, code and training data are all published under Apache-2.0.

How does it work?

The streaming half of Gander is built on MiniCPM-o 4.5 and uses a thinker-talker design: the Thinker decides what happens in the conversation, and a detached Talker renders speech, so the model can be cut off mid-sentence without losing its place. Input arrives in one-second chunks of video, speech and text. An orchestration runtime routes state to a pluggable back brain — the default provider is Codex — which runs tasks asynchronously and reports back.

Why does it matter?

Most voice assistants that also run tasks bolt an agent onto a speech pipeline, which is why they go silent while they work. Splitting the two lets Gander answer, ask a follow-up or take an interruption while the task continues underneath. Because the weights are Apache-2.0 and the whole thing serves locally, anyone building a voice interface can study or change that split rather than accept a vendor's version of it.

Who is it for?

speech and voice-agent researchers, teams building real-time assistants

Frequently asked questions

What hardware and setup does Gander need to run?: Gander ships as a local server. You clone the repository, build the conda environment from the included environment.yml, download the Gander-Omni/Gander checkpoints from Hugging Face into checkpoints/Gander, then run ./scripts/serve.sh and open http://127.0.0.1:8000. The stack also pulls faster-whisper-large-v3 for speech recognition alongside the Thinker and Talker weights.
How does Gander score against other speech models?: On SpokenQA, Gander reaches 75.60% on Llama Questions and 59.30% on Web Questions. On Full-Duplex-Bench v3 it records 100% turn-taking accuracy with 8% premature interruptions, and it scores 78.53% on the Daily-Omni multimodal understanding benchmark. The technical report on arXiv carries the full comparison tables.
Can I swap out the reasoning agent behind Gander?: Yes. Gander's back brain is pluggable — Codex is the default provider, but the orchestration runtime is what connects the streaming front end to whatever agent handles long-horizon work. That separation is the point of the Cerebellum-Brain design, and it is why the conversational half stays responsive while the reasoning half takes its time.
Is Gander free to use commercially?: Gander is released under the Apache License 2.0, which permits commercial use, modification and redistribution with attribution. Note that Gander is derived from the MiniCPM-o 4.5 base model and relies on faster-whisper-large-v3 for transcription, so anyone shipping it commercially should check those components' own terms as well.

Try it

hf download Gander-Omni/Gander --local-dir checkpoints/Gander && ./scripts/serve.sh