Quantized MiniMax H3 Turbo generates 5s video in ~1 minute on RTX 4060 Ti
aziib · reddit · 2026-09-02
Reddit user aziib tested the community-made minimax-h3-fused-turbo-int8-convrot quantized video model: only 4 sampling steps and about 1 minute to generate a 5-second, 0.4MP video on an RTX 4060 Ti 16GB, using sage attention and triton for speedup, with manual sigmas explored for further gains. Weights are on HuggingFace and a ComfyUI workflow is on Civitai, making it a complete low-VRAM setup for consumer GPUs.
More from Infra
- Can four M5 Ultra Mac Studios serve 100 users running Qwen locally? — Short_Report5589 · 2026-09-02
- Startup says OpenAI and Anthropic ignore TPM quota increase requests for months — Ok_Philosophy_4031 · 2026-09-02
- Superlinked open-sources Sie, an inference server for all agent-facing retrieval models — superlinked · 2026-09-02
- DLSS5 video enhancer arrives on Linux via open-source ComfyUI node — Euphoric-Let-5130 · 2026-09-02
- ComfyUI gains official AMD ROCm support on Windows for image and video generation — SeparateIntern5655 · 2026-09-02
- MLSecOps framework maps 10 security pillars for production ML systems — goyalshaliniuk · 2026-09-02