MiniMax's open-weight H3 video model is now fast enough that outside developers are building semi-real-time character chat apps on top of it.
The AMW Read
A community-built interactive video-chat demo on MiniMax's already-tracked open-weight H3 model extends a known speedup pattern rather than introducing new capability or a new debate.
MiniMax's open-weight H3 video model is now fast enough that outside developers are building semi-real-time character chat apps on top of it.
A Japanese developer built "FastH3 Date Chat," an app that lets a user converse with an AI character entirely through generated video instead of text or voice. The setup runs on two workstations: an RTX 4090 running MiniMax's accelerated "FastH3" variant for video generation, and an RTX A6000 running Google's local Gemma 4 12B model for dialogue. The user picks a response option or types free text, the LLM produces a reply in about 5 seconds, and FastH3 renders a matching 320x320 video clip in about 9 seconds, for a roughly 14-second round trip per exchange.
This extends a run of community-driven speedups around MiniMax's open-weight H3 release, following an earlier community extension that unlocked longer local clips via a Context Loop technique on an RTX 4090, and fal.ai's hosted H3 Max deployment, which cut 5-second clip generation to roughly 3 seconds. Each step lowers the latency floor until video generation starts to function as an interaction surface rather than a batch-rendering tool. MiniMax did not build this app — outside developers did, which is the payoff open weights are designed to produce.
For builders, the near-term opportunity is less about beating MiniMax on model quality and more about the orchestration layer: pairing a fast video model with a cheap local LLM to hit interactive latency, exactly as this demo does. For investors tracking generative video, the signal is that inference speed — not just output fidelity — is becoming a competitive axis, and that community adoption of an open-weight release can move faster than any single vendor's product roadmap.
