Gander is an open-source native duplex interaction model built around a Cerebellum-Brain architecture: a streaming Cerebellum handles realtime audio-visual perception and conversation while a separate Brain executes long-horizon agentic work, letting a user keep talking, interrupting, or redirecting a task while it runs. The Cerebellum, implemented in the minicpm_ft module as a fine-tune of MiniCPM-o 4.5, flattens video, audio, and optional task context into one-second causal units and predicts a control token (listen, speak, interrupt, or tool) before generating any content; a detached Talker then renders S3 speech tokens through a causal flow-matching decoder conditioned on the Thinker's hidden states, so speech generation does not block perception. The agent orchestration runtime, gander_runtime, turns the Cerebellum's task vocabulary (task_start, task_send with main or fork, task_resolve) into a tracked execution lifecycle, fencing stale results after a user revises an objective. Long-horizon reasoning and tool use are handled by a pluggable Back Brain behind a registry-based provider interface; the only provider included in this release is Codex, which requires a separately installed and authenticated Codex CLI (or a compatible executable), so agentic execution depends on an external provider rather than being self-contained to Gander's own weights. Gander was trained on a 2.7M-example mixture spanning speech interaction, audio-visual interaction, agentic interaction, and robustness/negative examples, using a two-stage recipe: the Thinker's LLM and audio projection train first, then are frozen while only the Talker's TTS projection and decoder train. Self-reported results include 75.60/59.30 on SpokenQA's Llama Questions/Web Questions splits, 100% appropriate turn-taking with an 8.0% premature-interruption rate on Full-Duplex-Bench v3, and 49.62%/78.53% accuracy on WorldSense/Daily-Omni. The release profile needs Linux, Conda, and three physical GPUs (one each for Thinker, Talker, and managed ASR), plus the MiniCPM-o 4.5 base, matching Gander checkpoints, and faster-whisper large-v3 for ASR. Code and the Gander/MiniCPM-o 4.5 weights on Hugging Face are Apache-2.0 and ungated; the training dataset is listed as coming soon and has not been released yet.