MiniMax H3 Open Source Weights: You Don't Need a Supercomputer. You Might Need a Lawyer.

MiniMax H3’s downloadable video-generation weights are available through official Hugging Face and ComfyUI repositories, offering separate fl2va checkpoints for text- and image-to-video generation and ref2va checkpoints for reference-guided video generation. Although the full model ecosystem is large, inference can use pruned and quantized files that reduce a single-task setup to about 42.5 GB, because roughly 13 billion of the model’s 33 billion parameters are AdaLN-related branches that need not be loaded for inference. ComfyUI reports that the model can run with dynamic offloading on a 12 GB GPU, though this trades speed for RAM usage, and local generation is natively limited to a 768-pixel short edge, while marketed 2K output requires a separate regeneration process. Benchmark reporting places H3 particularly strongly in video editing while indicating that competing hosted models lead some text-to-video and image-to-video categories. The weights are downloadable and modifiable but are governed by a restrictive community license rather than an OSI-approved open-source license: commercial products must display “MiniMax H3,” organizations exceeding $20 million in annual revenue need authorization, use to improve other AI models is prohibited, and the standard license excludes the EU, UK, South Korea, and United States. A hosted API remains globally available, offers direct 2K generation at per-second pricing, and may be more practical for smaller workloads or users in excluded territories, whereas local deployment favors users seeking customization, privacy, or high-volume use.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.