Kijai's new MH3 VAE cuts VRAM use: 1MP 10s video now runs on an RTX 3060 12GB
Chiduk99 · reddit · 2026-09-18
Kijai updated the VAE for MH3 (MiniMax-H3) three days ago, cutting VRAM usage with no quality degradation.
Real-world results on an RTX 3060 12GB + 16GB RAM:
- The old VAE maxed out at 0.8MP for 10 seconds;
- With the new VAE, the same machine can now generate 1MP at 10 seconds, and 0.7MP at 15 seconds;
- The post includes a sample R2V video made with 3 image references;
- The int8 convrot VAE is available on Hugging Face (Comfy-Org/MiniMax-H3).
More from Infra
- Payments firms race to own AI inference: Stripe taps OpenRouter, Ramp enters the chain — xkonjin · 2026-09-18
- Qdrant wraps 4+ hour Vector Space Stream on vector search — recording now live — qdrant_engine · 2026-09-18
- Investor: QNX's microkernel is an undervalued moat for the agentic AI era — pdamodaran · 2026-09-18
- Inception CEO Stefano Ermon argues diffusion will beat autoregressive models on inference efficiency — No Priors · 2026-09-18
- Inception CEO Stefano Ermon: Diffusion Beats Autoregressive LLMs on Inference — No Priors · 2026-09-18
- Zalando details its Kubernetes-based internal agent platform with kagent and kro — bibryam · 2026-09-18