What the official release establishes
MiniMax publishes H3-Base in BF16 checkpoints and documents local deployment through SGLang, vLLM, and Diffusers. ComfyUI separately provides quantized workflow assets and a native template path.
- H3-Base FL2VA and Ref2VA are separate task checkpoints.
- The initial open-source inference path uses full attention.
- ComfyUI documents INT8 diffusion weights and an NVFP4/AWQ text encoder for its templates.
There is no official universal consumer-GPU minimum
Neither the upstream README nor the ComfyUI guide currently gives one minimum VRAM number that applies across duration, resolution, task variant, offloading, and framework. We therefore do not label 12 GB, 16 GB, or 24 GB as an official requirement.
- Treat third-party GPU reports as configuration-specific evidence.
- Check the exact workflow, precision, resolution, duration, and offload settings before comparing reports.
- A model loading successfully does not establish acceptable generation speed.
Choose the route before evaluating hardware
Start with the intended workflow. ComfyUI templates use different packaged assets from the upstream BF16 serving routes, so hardware claims from one path should not automatically be applied to another.