Unsloth packages MiniMax H3 for local GGUF inference

Unsloth has published H3 GGUF quantizations for stable-diffusion.cpp and Unsloth tooling. The release converts MiniMax H3’s multi-file PyTorch stack into quantized denoisers and a compatible Qwen3-VL encoder. Its smallest denoiser file is 6.22 GiB, making single-GPU generation practical when the encoder and other components run from system memory.

H3 folds several workflows together

MiniMax H3 accepts context across text, images, video and audio. MiniMax says the complete system can generate clips up to 15 seconds at 2K resolution with synchronized stereo audio. The open checkpoint covers the base generation path, while MiniMax’s hosted pipeline handles the official 2K workflow.

Earlier video systems commonly used separate checkpoints for text-to-video, image-to-video, first-and-last-frame conditioning, subject references, motion references and video editing. H3 expresses those relationships through natural-language instructions within one model family.

Use the camera movement from Video 1, make the character in Image 2 sing, and match the vocals to Audio 3.

A prompt can therefore coordinate the picture, dialogue, sound effects and music within the same generation process, preserving timing across the supplied references.

The download has four moving parts

Unsloth quantized the official checkpoint, including two denoiser variants and their shared Qwen3-VL 32B encoder. The denoiser iteratively converts latent noise into media; the encoder converts prompts and conditioning context into features that guide that process.