November 11, 05:39Pandaily

MiniMax releases open-weights H3-Image model for 16 MP image generation and editing

MiniMax has introduced the MiniMax-H3-Image model with open weights. This model is designed for text-to-image conversion and image editing tasks and is part of MiniMax's H3 family.

The release occurred late on November 10th, and the weights are available on the Hugging Face platform. According to MiniMax, this model is capable of generating an image with a resolution of approximately 16 megapixels, sized at 4096 by 4096 with a 1:1 aspect ratio, in under 10 seconds. The company claims this is about five times faster than the market average for such a resolution.

Model Features and Capabilities

According to the model card, one checkpoint file can perform both text-to-image conversion and edit one or more reference images, accepting up to 16 such references. The diffusion transformer has 33.1 billion parameters, uses the H3 architecture but without audio layers, and was distilled to six sampling steps without using a classifier-free guidance. The model's text encoder is a separate component—the MiniMax-H3 text encoder.

MiniMax indicates that output dimensions can reach 6272 by 2688 without using an upscaler. The BF16 format file weighs 68.65 GB, and the additional prompt extension model, named MiniMax-H3-Image-PE, is 17.9 GB.

The company states that the model's instruction following is the strongest among open-source image models, and it avoids the characteristic 'artificial intelligence' look. These statements are the company's own claims, and no independent comparisons were present in the original material.

The model card recommends using SGLang, vLLM-Omni, diffusers, and ComfyUI for local deployment, and also lists NVIDIA and AMD accelerators. The model is also available via API and MiniMax applications, and for the first week, the company is offering a 50% discount on API requests for 16MP and higher resolution images.

Licensing Terms

The license applied to the model is called the MiniMax-H3-Image Community License Agreement, and it is not a standard open-source license. This agreement is valid worldwide, except in the European Union, the United Kingdom, South Korea, and the United States, stipulating that use outside this territory is not sanctioned by the agreement. Furthermore, it requires obtaining separate written permission from MiniMax for users whose annual income exceeds 1 million US dollars, or for those who provide hosting services to third parties. Commercial products are also required to display the inscription 'MiniMax-H3-Image' in their interface.

This release followed an earlier release by MiniMax of open weights for video and audio based on the H3 model.

Similar stories

technode.comNovember 5, 09:01

Moonshot AI releases weights for Kimi K3.1 Pro model for independent deployment

Moonshot AI has published the weights of its flagship multimodal model, Kimi K3.1 Pro, on the Hugging Face platform, allowing users to deploy the model independently.

The model is built on a mixture-of-experts architecture and has a total of 2.8 trillion parameters, with approximately 104 billion parameters activated during each inference. Additionally, it supports a context window of about one million tokens.

The K3.1 Pro version is an evolution of the Kimi K3 model, incorporating advanced reinforcement learning methods for tasks related to coding, 3D object generation, financial research, and long-term agent execution.

The release of the weights followed the model's availability through the Kimi service's applications and API. It was also announced that Huawei provided inference support via Atlas A3 and Ascend 950 systems using vLLM Ascend and SGLang.

technode.comNovember 4, 12:15

Tencent releases Hy4 model with computer vision and control capabilities

Tencent has introduced its flagship artificial intelligence model, Hy4, which possesses visual understanding capabilities and the ability to interact with a computer. The model was launched on November 4th.

Hy4 is capable of analyzing charts and documents, reasoning based on multiple images, as well as controlling command-line tools and graphical interfaces. The training process covers applications such as Blender and CAD software, as well as tasks requiring simultaneous operation across multiple programs.

The Hy4 model has 770 billion parameters and is available in Tencent's own products as well as through third-party platforms, including OpenRouter. Soon, the company plans to publish the model weights and a dedicated system card.

The standard API pricing remains unchanged compared to the Hy4 preview: the cost is $0.834 per million input tokens and $2.501 per million output tokens. Additionally, a 50% discount is offered on Tencent Cloud TokenHub until December 31st.

Tencent has integrated Hy4 into the WorkBuddy and CodeBuddy products, providing a limited free usage quota from November 4th to November 15th. A more cost-effective version of the model, Hy4 Flash, is also planned. The company noted that development focused on office tasks, financial analysis, and game creation, with development teams assisting in training and evaluating the model based on real workflows.

pandaily.comOctober 29, 10:07

DiffuSpace releases Clover-35B-A3B model—a diffusion language model for agents and coding

The DiffuSpace team, which previously developed the diffusion language model Dream 7B, has released and made publicly available the Clover-35B-A3B-Preview model. This model is designed to perform tasks traditionally handled by conventional large language models, such as writing code, interacting with tools, browsing the web, and creating frontend interfaces.

Most existing language models operate autoregressively, meaning they predict tokens sequentially, where each new token depends on the previous ones. Diffusion language models use a different approach: they update multiple positions simultaneously and refine the result over several rounds, reconstructing hidden or noisy text into a coherent response. This method promises faster generation, although historically, diffusion models have shown unreliability in agent tasks, often leading to incorrect tool calls and repetition of failed steps.

Clover utilizes a block diffusion method. The output stream is divided into blocks that are generated sequentially, while positions within each block are denoised in parallel. The team reported that they achieved this by restructuring the Gated DeltaNet foundation for parallel block-level denoising and adding a trainable block size calibration. The model features a 'mixture of experts' architecture with a total of 35 billion parameters and approximately 3 billion active parameters per token, and it supports a thinking mode.

According to data published by the team, Clover achieved a score of 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. Regarding tool usage, the model reached 65.3 on the general tau-cubed benchmark and 67.9 on BFCL-v4. The team claims these metrics surpass those of similarly sized autoregressive models, including Qwen3.6-35B-A3B. It should be noted that all these figures are self-reported and have not yet undergone independent verification.

The primary advantage is speed. DiffuSpace reports a speed of 515 tokens per second when performing agent tasks on a single NVIDIA H200 chip, compared to 365 tokens per second using the DFlash2 speculative decoding setting, and a general speed of 680 tokens per second on a single GB300. In a web agent test scenario covering 49 tasks, Clover's median execution time was 1.52 seconds.

The model weights are available on Hugging Face, GitHub, and ModelScope under the Apache 2.0 license. DiffuSpace also offers trial access through its website and reports developing tooling for agents and a desktop client for coding. Since Clover is designated as a preview version, its reliability in real-world long agent workflows will require external testing. Nevertheless, it is one of the largest open diffusion language models specifically focused on coding and agent tasks.