MiniMax releases H3-Image open weights for 4K generation in under 10 seconds
MiniMax released H3-Image on November 10, 2026, and its research blog post is dated November 11. The model is an open-weight image generator built on H3, the company's video model. It covers text-to-image generation, reference-based generation and image editing. MiniMax says it can produce a 4096 by 4096 image in under 10 seconds, and it claims top prompt adherence among open-weight models. The weights are hosted on Hugging Face, and the model also runs through the MiniMax API, MiniMax Code and Hailuo AI.
The company attributes its speed and quality gains to changes across the full post-training pipeline, including data, reinforcement learning, fewer sampling steps and deployment work. A GitHub README describes a 33.1 billion parameter diffusion transformer that renders roughly 16 megapixel images at any aspect ratio in six sampling steps, without a separate upscaler. Per TechCity Authority's reading of the launch announcement, the company offers 50% off generations of 16 megapixels or more during launch week.
Coverage differs on some details. Pasquale Pillitteri's report gives a 10 to 20 second generation range, while MiniMax says under 10 seconds. Crypto Briefing reports that the community license excludes users in the US, EU, UK and South Korea. The license text itself was not checked for this brief, so readers should review it directly before relying on that claim.
Why it matters
For teams building products, the main change is latency. A 4K result in about ten seconds, if independent tests confirm it, makes high-resolution generation workable inside interactive features rather than only in batch jobs. Open weights add a second route: self-hosting on owned or rented GPUs. A checkpoint of roughly 69 GB plus its text encoder makes infrastructure planning a real cost item. The speed figures are vendor claims, so latency budgets should be tested against real prompts and hardware.
For investors and sellers, the release adds pricing pressure on closed image APIs that charge per image, and it raises the bar for editing and reference handling. License terms may matter as much as benchmark scores. If the reported territorial exclusions apply, distribution plans, enterprise contracts and compliance reviews in those markets could change. Due diligence on image-generation vendors should now cover weights licensing, not just output quality.
Key points
- 1Native output reaches 4096 by 4096 pixels, about 16 megapixels, without a separate upscaler pass.
- 2MiniMax claims under 10 seconds per 4K image, while one independent report cites 10 to 20 seconds.
- 3The diffusion transformer has 33.1 billion parameters and weighs about 68.65 GB in BF16 precision.
- 4One checkpoint covers text-to-image, reference-based generation and editing, with up to 16 reference images.
- 5
