Companies in the US, EU, UK and South Korea need a separate license from MiniMax to use the open weights, and larger users elsewhere need its written approval.
MiniMax has published the weights of H3-Image, an image generation and editing model that the Hong Kong-listed company says can produce a 4,096-by-4,096-pixel picture in under 10 seconds. The model takes text prompts or up to 16 reference images and is available on Hugging Face as well as through MiniMax’s paid API, TechNode reported on November 11. GPU maker Moore Threads said its chips supported the model from the day the weights went out.
The license attached to the weights limits where and how they can be used. Under the MiniMax-H3-Image Community License, use is permitted worldwide except in the European Union, the United Kingdom, South Korea and the United States, where parties must ask MiniMax for a separate license, and larger commercial users need MiniMax’s written approval. The speed claim also comes without a stated hardware setup, so outside developers will need to reproduce it.
The release extends the H3 family that MiniMax launched in July. EastFrontier reported then that MiniMax’s H3 multimodal model ranked first in Artificial Analysis’s video editing category and would be released under MiniMax’s community license. H3-Image builds on H3’s pretraining for still images.
A 33-billion-parameter model distilled to six steps
The Hugging Face model card gives the technical detail MiniMax’s announcement leaves out. The core image generator is a diffusion transformer, the kind of network that builds an image by refining noise step by step, with 33.1 billion parameters across 50 layers. It reads prompts through the text encoder of MiniMax H3, which the card describes as the full Qwen3-VL encoder, and it has been distilled to six sampling steps, with no classifier-free guidance pass. Fewer steps mean less computing per image.
Output runs at about 16 megapixels across aspect ratios, from 4,096 by 4,096 for a square image to 6,272 by 2,688 for a 21:9 frame. The main model weights take 68.65 GB in BF16 precision, and the card lists Nvidia’s H200 and RTX PRO 6000 Blackwell and AMD’s Instinct MI350X among supported accelerators. It suggests spreading work across several GPUs, for example four, for lower latency.
MiniMax’s blog post says 4K generation is up to five times faster than mainstream alternatives, without naming them or the hardware used. It calls the model’s prompt adherence state of the art among open-weight models, though it cites no benchmark or score, and neither does the model card. Through the API, MiniMax prices the model at 50% to 70% of comparable mainstream offerings and is offering half price on 4K and larger requests during launch week. The model is also available in MiniMax Code and Hailuo AI.
The license keeps commercial control with MiniMax
The license grants a royalty-free right to use, modify and distribute the model within its permitted territory. MiniMax’s license requires written authorization from the company once a user’s annualized revenue exceeds US$1 million or it offers the model as a hosted service. Commercial products built on it must display the MiniMax-H3-Image name prominently in their user interface.
Users also may not use the model or its outputs to improve other AI models, apart from H3-Image and its derivatives. MiniMax claims no rights over the images users generate, and the agreement is governed by Hong Kong law. The license itself does not address whether the same territorial limits apply to MiniMax’s paid API.
Domestic chip support arrived on day one
Moore Threads said its MTT S5000 accelerator for AI training and inference, together with its MUSA software stack, had been adapted to H3-Image on the day the weights were released, IT Home reported. The company reused existing components including SGLang, MATE and MCCL, having already built core operations such as attention and matrix multiplication for MiniMax H3. Moore Threads pointed to the heavy attention load of diffusion models with many reference images as a workload suited to the S5000’s computing density.
Day-zero support from a Chinese GPU maker gives MiniMax a domestic deployment route alongside the Nvidia and AMD hardware listed on its model card. The next marker is independent testing: whether developers running the weights on H200s or S5000s match the under-10-second claim at full 4K resolution, and how H3-Image ranks once leaderboards such as Artificial Analysis add it.