On November 10, MiniMax Xiyu Technology released the open-weight image generation model MiniMax H3-Image and opened its weights. The model is derived from H3 and supports text-to-image, image reference, and image editing. It can generate 4K images at 4096×4096 pixels, about 16MP, in 10 seconds—5x the speed of mainstream offerings on the market. The weights are now available on Hugging Face and are offered through the MiniMax Open Platform, MiniMax Code, and Hailuo.
H3-Image strengthens the post-training pipeline on top of H3's pretraining capabilities, covering data, reinforcement learning, step compression, PE, and deployment optimization, while balancing speed, image quality, and instruction following. It uses rich detail and realism to reduce the common "AI feel" and offers the strongest instruction-following capability among open-source models, supporting up to 16 reference images. Compared with mainstream 1MP or 4MP outputs, the 16MP specification has a significant detail advantage, making high-resolution everyday creation possible.
Official P50 median latency benchmarks show that 4K text-to-image takes 16.0 seconds in Standard mode and 10.3 seconds in Highspeed mode, compared with 50.0 seconds and 51.0 seconds for Nano Banana 2.1 at the same specification; for image editing with a single 4K image input, the times are 18.2 seconds and 12.5 seconds, compared with 48.0 seconds and 54.0 seconds for other models at the same specification.
Overall, Standard mode is 1.5–2x the speed of mainstream tools on the market, while Highspeed mode is 3–4x faster. Pricing is only 50%–70% of same-tier mainstream tools. For 4K (16MP) text-to-image, it costs ¥0.2261 per image (original price ¥0.4522), and ¥0.4275 for the Highspeed version (original price ¥0.8550). In the first week after launch, generation requests at 16MP and above get 50% off the official price.
MiniMax said it will continue the same open-weight strategy as H3 and hopes H3-Image will be the starting point for narrowing the gap between open-source models and top closed-source models. This simultaneous emphasis on speed, unit price, and open weights reflects that competition in image generation is shifting from image-quality battles to efficiency and cost in high-resolution scenarios.