H3 Turbo 8-step 768p, baked W4A8
Community conversions for ComfyUI. Each file combines a pruned MiniMax H3 model with its matching LightX2V 8-step v1.0 768p adapter at strength 1.0. The adapter is merged before quantization.
| File | Mode | Quantization |
|---|---|---|
minimax_h3_fl2va_pruned_turbo_8step_v1.0_768p_w4a8_baked.safetensors | Text and first/last frame to video/audio | 200 W4A8 block layers; BF16 refiner |
minimax_h3_ref2va_pruned_turbo_8step_v1.0_768p_w4a8_baked.safetensors | References to video/audio | 200 W4A8 block layers; 8 INT8 ConvRot refiner layers |
Both retain the remaining tensor types from Kijai's corresponding mixed W4A8 file. They use codebooks, groups of 16 weights, ConvRot groups of 256 and FP8 group scales. They need ComfyUI support for comfy-kitchen W4A8.
Use
Place the desired file in ComfyUI/models/diffusion_models. Load it with the
normal diffusion model loader, in an H3 workflow for the corresponding mode.
Remove the LightX2V Turbo LoRA loader: the adapter is already in the file.
Recommended adapter settings: 8 steps, Euler, simple schedule, video shift 6, audio shift 3, CFG 1, Native 768p-class canvas. The text encoder and the audio and video VAEs are separate files. Normal, SolAttn and VedaSparse remain workflow choices. The baked adapter strength is fixed at 1.
These are Turbo models. Use the original model files for Base generation or for an adjustable Turbo adapter strength.
Generation checks and speed
Test hardware: A100-SXM4-40GB, 6 physical CPU cores / 12 threads, 83.5 GiB RAM.
ComfyUI 46493d824fd7eac25a741e505159e4c9e0ef0ebd, comfy-kitchen 0.2.37,
PyTorch 2.13.0+cu130. No upscaler or sparse attention was used.
All clips: 1344×768, 124 frames (5.17 s), 8 steps, res_multistep, simple
schedule, shifts 6/3, CFG 1, seed 42. FL2VA used text; Ref2VA used one image.
These sampler settings match the existing benchmark, rather than the Euler
recommendation above. Downloads are outside the generation timer.
| Test | Original model + adapter | Baked model |
|---|---|---|
| FL2VA, original first | 153.0 s | 186.3 s |
| FL2VA, baked first | 162.3 s | 154.2 s |
| Ref2VA, original first | 201.2 s | 176.6 s |
The saved L4 FL2VA baseline was 376 s with the same input and software. The original model was 2.3–2.5 times faster on the A100. These few runs do not establish a consistent speed gain from baking. Cache and initialization costs varied. ComfyUI recorded zero runtime adapter patches for the baked files.
The first FL2VA pair and the Ref2VA pair passed full video decode, frame count, resolution, audio-stream and checksum checks. This is a functional check with anime content, not a broad quality evaluation.
Sources and method
- Pruned BF16:
Comfy-Org/MiniMax-H3, revisione5eb578a89295337b8ff433a035929ce0279e0b6. - Matching
*_8step_v1.0_768p_comfyui_bf16.safetensorsadapters:lightx2v/Minimax-h3-Turbo, revision3ec17a324ced54151364f24f8b5fb6bf7e26414f. - Tensor layout and quantization settings:
Kijai/MiniMax-H3-experimental, revisione042fe480f58806578713532b8ae4e3d47d1bd63.
bake_h3.py checks all 208 adapter targets. It computes
W + (alpha / rank) * B @ A in FP32, casts to the source tensor type, and
quantizes once with nearest rounding. It does not start from the already
quantized diffusion model. The output tensor names, shapes and types must
match the reference layout before a file is saved.
The bake reports contain exact revisions, file hashes, quantization errors, package versions and hardware. Generation tests confirm that the files load and produce video with audio. A different output at the same seed is expected; the quantized weights differ from runtime adapter merging. Baking does not guarantee identical output across software or GPU versions, or better quality for every prompt.
License
These files derive from MiniMax H3 and remain subject to the MiniMax H3
Community License Agreement in LICENSE-MiniMax-H3. The LightX2V adapter
license is provided separately in LICENSE-LightX2V (Apache 2.0).