Minimax_h3_hybrid

About

What Minimax_h3_hybrid is

MiniMax H3 Hybrid NVFP4 for NVIDIA Blackwell Two experimental mixed-precision quantizations of the MiniMax H3 Ref2VA Pruned diffusion model, optimized and tested locally on an NVIDIA GeForce RTX 5070 Ti 16 GB (Blackwell, SM 12.0). This is an independent community optimization by abakanai. It is not an official MiniMax or ComfyUI release. The checkpoints are inference-only single-file diffusion weights for ComfyUI; they are not LoRAs and do not include the text encoder or VAEs. Checkpoints The attention-output and FFN-output projections remain INT8 ConvRot in both variants. This preserves the fused SwiGLU path and avoids spending precision where native NVFP4 did not provide a measured speed benefit. Generated examples All media below are AI-generated. The paired videos in each comparison use the same reference image, prompt, seed, resolution, duration, sampler, scheduler, and step count. Matched-seed comparison - Conservative hybrid video with audio - Fast hybrid video with audio A/B generation settings In this single-seed qualitative check, both variants preserved the character's face and glasses across the clip. The conservative variant held finger shapes slightly more consistently, while the fast variant produced a very similar result with somewhat larger changes in background composition and framing. This is one visual sample, not a statistically meaningful perceptual evaluation. 1376 × 768 simple follow-up The two hybrid checkpoints were also tested at 1376 × 768 using the same reference, prompt, seed, 124-frame duration, 20 steps, and resmultistep sampler. The scheduler was changed to simple for both matched runs. - Fast 14.03 GiB high-resolution video with audio - Conservative 16.38 GiB high-resolution video with audio Both runs completed on the RTX 5070 Ti 16 GB through ComfyUI DynamicVRAM and produced 124-frame, 24 fps MP4 files with stereo AAC audio at 32 kHz. The fast checkpoint finished about 5% sooner in these sequential single runs. This is a practical observation rather than a controlled end-to-end benchmark because model and cache state were not fully reset between jobs. Research summary Test environment - NVIDIA GeForce RTX 5070 Ti, 16 GB - Black

Summarised from the published model card. Read the full card on the HuggingFace links below.

Specifications

What it is

How it works

How video models work

Variants

Sizes & precisions

Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.

Using it via the API

Call it like any OpenAI endpoint

Once AxForge deploys minimax-h3-hybrid for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (minimax-h3-hybrid below is illustrative; you get the exact model name on deployment.)

$ curl -sS https://api.axforge.ai/v1/videos/generations \
  -H "Authorization: Bearer $AXFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"minimax-h3-hybrid","prompt":"a drone shot over a forest"}'

Details

Languages, data & research

Licence

Commercial licence needed

Sources

Weights & code

Explore

More video models