Overview
Gander is an end-to-end omni-interaction agent from Tencent's Hunyuan Speech Team, released with its technical report on 9 September 2026. Where a conventional multimodal assistant works in turns — you finish speaking, it answers — Gander continuously receives streaming input across video, speech and text at the same time, and unifies omni perception, real-time interaction and agentic capability in a single model.
The architecture is what the report calls a Cerebellum-Brain collaborative framework, implemented as a streaming Thinker-Talker design over a chunk-level token stream. The fast half handles real-time audio and video and produces speech; the slower half carries long-horizon reasoning. That split is what allows genuine full-duplex behaviour: the user can interrupt mid-answer, and the model can volunteer feedback without waiting to be prompted. The published operating envelope is a one-second temporal unit, BF16 weights, 16 kHz audio input and 24 kHz speech output, with a budget of up to 8 lexical tokens per speaking unit for the Thinker and 50 S3 tokens per speaking unit for the Talker.
Weights, code and the project page are all published openly. The model card gives the base model as openbmb/MiniCPM-o-4_5 and inherits its Apache 2.0 licence, and the GitHub repository carries the same licence. Tencent positions it for both everyday conversation and workflow-oriented agent scenarios — asynchronous, long-horizon execution alongside live voice collaboration — rather than as a general text flagship, which is what the Hunyuan line covers.
| Released | 2026-09-09 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Architecture | Cerebellum-Brain collaborative framework with a streaming Thinker-Talker design over a chunk-level token stream, built on openbmb/MiniCPM-o-4_5 |
| Modalities | Text, Audio, Vision, Video, Speech output |
| Status | Research release — the technical report and weights were published together; the project page describes it as a research project of Tencent's Hunyuan Speech Team. |
Benchmarks
- Daily-Omni (audio-video understanding)78.53%
- SpokenQA — Llama Questions75.6%
- SpokenQA — Web Questions59.3%
- WorldSense (audio-video)49.62%
- VoiceBench — SD-QA46.84%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Full-duplex spoken interaction with published turn-taking measurements: 100.0% timely response rate and an 8.0% premature-interruption rate on Full-Duplex-Bench v3
- Simultaneous streaming perception of video, speech and text rather than turn-by-turn multimodal input
- Apache 2.0 open weights, code and data, free for commercial and self-hosted use
- Strong audio-video understanding for its class — 78.53% on Daily-Omni
- Cerebellum-Brain split lets real-time speech continue while long-horizon reasoning runs behind it
- Documented streaming envelope (1-second temporal unit, 16 kHz in / 24 kHz out, per-unit token budgets) makes latency behaviour predictable
Best for
- Voice assistants that need to be interruptible mid-sentence rather than waiting for a turn to end
- Live screen-and-voice collaboration where the model watches a video stream while you talk over it
- Agentic workflows that mix spoken instruction with asynchronous, long-running execution
- Self-hosted omni-modal research on top of an Apache-2.0 checkpoint with published benchmarks
How to access
| Provider | Model ID |
|---|---|
| Hugging Face (self-hosted) ↗ | Gander-Omni/Gander |
FAQ
Is Gander open source?
Yes. The GitHub repository is Apache-2.0, and the Hugging Face model card carries the Apache 2.0 licence inherited from its base model, openbmb/MiniCPM-o-4_5. Weights, code and training data were published together with the technical report.
How many parameters does Gander have?
Tencent has not published a parameter count on the model card, the project page or the technical report abstract, so we do not list one. What is documented is the precision (BF16) and the base model it is built on, MiniCPM-o 4.5.
What does full-duplex mean here, and how well does Gander do it?
Full-duplex means the model listens and speaks at the same time, so you can cut in mid-answer and it can speak up unprompted. On Full-Duplex-Bench v3 across 100 scenarios, the project page reports a Pass@1 of 0.400, a 100.0% timely response rate and an 8.0% premature-interruption rate.
What inputs and outputs does Gander handle?
The model card lists text, audio, image and video frames as inputs, with text and speech as outputs. Audio comes in at 16 kHz and speech is generated at 24 kHz, on a one-second temporal unit.
How does Gander relate to Tencent's Hunyuan models?
It is a separate line. Hunyuan Hy3 and Hy4 preview are large text Mixture-of-Experts flagships; Gander comes from Tencent's Hunyuan Speech Team as a research release focused on streaming audio-visual interaction and agentic task orchestration.