MiniMax (00100) Open-Sources H3-Image Model, With Major Global Chipmakers On Board on Day One

On the evening of November 10, MiniMax (00100) released the image generation model MiniMax H3-Image and simultaneously opened the model weights. Derived from MiniMax H3, the model supports text-to-image, image reference, and image editing. Building on the broad image capabilities endowed by H3 pre-training, H3-Image further strengthened the entire post-training pipeline. Improvements in data, reinforcement learning, step compression, PE, deployment optimization, and other aspects give H3-Image competitive strength in speed, image quality, and instruction following simultaneously. According to the introduction, MiniMax H3-Image can generate a high-specification 40964096-pixel image of approximately 16MP (4K) within 10 seconds, at a speed up to 5 times that of mainstream offerings on the market. In terms of image quality, the model provides rich visual detail and extremely strong realism, further narrowing the gap between AI-generated images and the real visual world. In addition, H3-Image possesses top-tier instruction-following capability among open-source models and supports up to 16 reference images. MiniMax also offers H3-Image image generation services through the open platform, MiniMaxCode, and MiniMax Hailuo. The generation speed of H3-Image in standard mode is approximately 1.5 to 2 times that of mainstream tools on the market, and the Highspeed mode can reach 3 to 4 times, with service pricing at approximately 50% to 70% of mainstream tools in the same tier. During the first week of release, MiniMax is offering a 50% discount on official pricing for generation requests at 16MP and above. Three months after the release of MiniMax H3, MiniMax extended its open-weight strategy from video generation to image generation. In terms of model architecture, H3-Image removed the audio branch from H3's original architecture and was optimized for image generation and editing. Technical materials published on HuggingFace show that H3-Image adopts a 33.1-billion-parameter dense diffusion Transformer (DiT) with 50 layers, continues to use MiniMax H3's Qwen3-VL text encoder, and combines an image VAE to process reference images and generated images. Enterprises and developers can explore new tools and creative methods based on the model weights, or directly use high-quality generation services through official channels such as the MiniMax open platform. Through the official H3-Image API, individual developers can integrate it into daily workflows within agents such as MiniMax Code, WorkBuddy, and Codex, making image generation a foundational capability callable by various Agent frameworks. The release of H3-Image raises the capability ceiling of open-source image models to a new dimension: whether they can accurately understand complex creative requirements and whether they can translate those requirements into aesthetically pleasing and realistic images. An image containing text and multi-layered information must clearly convey content and relationships while also making composition, lighting, and detail serve the creator's intended expression. H3-Image combines instruction following, multi-image reference, and high-resolution generation, continuing MiniMax's multimodal integration approach and offering new possibilities for open-source models to narrow the gap with top closed-source models in complex visual creation. On the first day of open-sourcing, MiniMax partnered with global chip companies including Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren, Hygon, MetaX, Iluvatar CoreX, NVIDIA, AMD, Intel, and Tenstorrent, as well as multiple mainstream inference and cloud service platforms, to advance Day0 integration work. MiniMax stated, "We hope H3-Image can bring convenience to more people's daily lives and work, and we will continue to build stronger tools to help realize inclusive AGI."