The practical barrier for many creators is not model capability but hardware: large multimodal encoders often require 32–80 GB GPUs. This re-quantization brings the uncensored MiniMax-H3 text encoder down to 15.7 GB while preserving output parity with the upstream INT8 ConvRot build — making single‑GPU (16 GB) video generation feasible without changing downstream pipelines.
What Sets It Apart
- Fits on 16 GB GPUs: the baked file is 15.7 GB and was measured to generate 6 s of 480×864 vertical video with peak VRAM ~9.9 GB on an RTX PRO 2000 Blackwell (16 GB). This makes MiniMax-H3 text encoding viable on consumer/prosumer cards.
- ConvRot-aware re-quantization: upstream weights were stored rotated (W_stored = W @ H^T per 256-group). The bake explicitly un-rotates weights (multiply by Hadamard H) before NVFP4 quantization; skipping this yields unrelated or meaningless conditioning despite no runtime errors.
- Mixed-precision strategy: most linear layers are NVFP4 (TensorCoreNVFP4Layout, group size 16) while the large embed_tokens layer remains INT8 to avoid OOM during baking; ComfyUI per-layer metadata supports this mix transparently.
- Practical constraints documented: inherits upstream INT8 rounding choices, requires NVFP4-capable hardware (e.g., NVIDIA Blackwell with NVFP4 support), and is distributed as a drop-in CLIPLoader replacement for existing MiniMax-H3 ComfyUI workflows.
Who it's for and trade-offs
Great fit if you need to run MiniMax-H3 video generation on a single 16 GB GPU, want a drop-in replacement for ComfyUI CLIPLoader, and accept uncensored (Heretic) encoder variants under the MiniMax H3 Community License. Look elsewhere if you lack NVFP4 hardware (hardware support is required for NVFP4 acceleration), require a bake from BF16 (this build re-quantizes from INT8 and thus inherits its rounding), or cannot use uncensored weights for policy/compliance reasons. Also note that the diffusion model and VAEs from Comfy-Org/ MiniMax-H3 are still required for full generation.
Where It Fits
This artifact is a compatibility-and-efficiency engineering result rather than a new architecture: it trades higher-precision storage for a quantized layout tailored to NVFP4 Tensor Cores, preserving behavioral parity with the upstream INT8-ConvRot encoder while enabling practical single‑GPU workflows.
