MiniMax has released the model weights for "MiniMax H3," a model that generates and edits video and audio simultaneously. As of August 4, 2026, H3 ranks first in Artificial Analysis's video editing leaderboard with audio. It also ranks second in text-to-video generation and third in image-to-video generation. The significance of being able to take a high-performance video model outside of an API is substantial, but the hosted 2K version that took first place and the self-contained public version do not share the same configuration.
What was released is "H3-Base," the core model that generates 768p video and stereo audio. "H3-Context-IR," which converts complex reference material into generation instructions, and "H3-Regenerate-2K," which reconstructs output at 2K, remain on MiniMax's servers. Cross-referencing the rankings with what was made public reveals both the fact that H3 has narrowed the performance gap with proprietary models and the gap that still cannot be closed in a local environment.
How to Interpret the No. 1 Ranking in Video Editing
In Artificial Analysis's Video Editing category with audio, H3 ranked first with an Elo of 1,130 across 8,211 evaluation samples. Second place, Gemini Omni Flash, scored 1,122 with 10,847 samples. Although the 95% confidence intervals for both models overlap at ±6, H3 holds the top position in the site's displayed ranking.
Even for the same H3, the ranking changes depending on the input method. In Text to Video with audio, it ranked second with an Elo of 1,239 across 6,031 samples, while first-place Gemini Omni Flash scored 1,244. H3's displayed rank range is 1st to 2nd. In Image to Video with audio, it ranked third with an Elo of 1,187 across 5,382 samples, but the rank range spans 1st to 3rd place. There is no basis for combining these three rankings into a claim that it is "the best across all video generation."
In the evaluation, two videos generated from the same instruction are shown to users with the model names hidden, and users choose their preference. Artificial Analysis aggregates the pairwise votes using the Bradley-Terry method and converts them into an Elo-like metric. With audio versus without audio, video editing versus text generation, and image-based generation each constitute separate populations, so scores cannot be directly compared across categories. Rankings are recalculated hourly, so the figures cited here are also a snapshot from August 4.
Even so, the top ranking in video editing carries technical significance. H3 takes existing video and text instructions as input, changing the motion and characters in the footage while also outputting audio together with it. Compared to a process where only the visuals are created and dialogue or sound effects are layered on afterward using a separate model, it is easier to align on-screen events with audio timing within a single generation process. A preference-based ranking does not guarantee the accuracy of the editing content, but it does indicate that the integrated output was preferred by users.
What Was Released Is the Middle Stage of a Three-Stage Pipeline
The complete H3 system operates through three stages: H3-Context-IR, H3-Base, and H3-Regenerate-2K. The first stage, Context-IR, interprets relationships spanning text, images, video, and audio, converting them into structured instructions that the generation model can easily handle. The final stage, Regenerate-2K, feeds the 768p result and the original reference material back into H3, reconstructing 2K video using context rather than simply upscaling details through inference.
H3-Base, published on Hugging Face, is split into two BF16 checkpoints. The FL2VA version handles text-only generation as well as generation with specified first and last frames. The Ref2VA version is designed for referencing images, video, and audio, carrying over people, motion, camera work, and voice. Both output video together with 32kHz stereo audio simultaneously.
Meanwhile, Context-IR and Regenerate-2K are not included in the public package. MiniMax explains that the former relies on multiple hosted models and services, and states that the latter will be released once it is ready. Artificial Analysis also explicitly noted that it evaluated H3 at the 2K tier. Therefore, the top-ranking score is not the result of running only the published H3-Base locally, but an evaluation of the complete system MiniMax provides.
ComfyUI version 0.30.0 and later includes a standard workflow for H3, allowing users to create video locally from text, images, and multiple reference materials. However, the native local output has a short side of 768 pixels, with a maximum of 768×1,344 pixels. The official procedure for reaching 2K combines the local H3-Base with two of MiinMax's APIs. While the public weights expand freedom for editing and fine-tuning, they do not fully bring the same pipeline used in the ranking onto local hardware.
A 33-Billion-Parameter Transformer Predicts Video and Audio Simultaneously
At the core of H3-Base is a dense, single-stream Transformer with 33 billion parameters. Rather than passing video and audio to separate dedicated models, the same Transformer predicts latent representations for both. On the input side, H3-Encoder, which uses all the trained weights of Qwen3-VL-32B, encodes text and visual information. This means describing "H3 as a whole" as having 33 billion parameters is a count that excludes the Encoder and other components.
Output ranges from 4 to 15 seconds at 24fps, with 32kHz stereo audio. For reference-based generation, it accepts up to 9 images, 3 videos, and 3 audio clips. Video and audio each have a combined limit of 15 seconds, and mixed inputs are capped at a total of 12 files. Audio alone cannot be provided as a reference; it must be accompanied by an image or video.
Design choices to reduce computational load are also included. The video VAE compresses space by a factor of 16 and time by a factor of 4, and combined with subsequent splitting, the spatial reduction ratio entering the Transformer is 1/32. MiniMax states that the new VAE has increased the effective sequence length fourfold. However, the initial public release only supports full-attention inference, with the sparse attention implementation introduced during training scheduled for release at a later date. Local execution speed and memory requirements may change with future implementations and quantization.
Behind H3's strength in video editing lies a design that represents the relationship between reference and edit in natural language, rather than lining up dedicated models for each use case. For instance, a character is carried over from an image, motion follows a video, and voice references an audio clip—three roles processed as a single unified context. However, in the local version, users must compensate for the input organization that Context-IR handled. With the release of the weights, prompt processing has become a development target in addition to the model itself.
2K Costs $0.13 per Second, Public Weights Come with Regional Restrictions
MiniMax's API pricing is $0.13 per second for 2K output and $0.08 per second for 768p. For a 10-second clip, that comes to $1.30 and $0.80 respectively; for 15 seconds, $1.95 and $1.20. Using only the 2K regeneration costs $0.05 per second. The first 5 reference images are free, with a $0.04 charge per image starting from the 6th. Video input and Context-IR also carry separate usage-based fees, so for production work involving repeated trial and error before completion, the total cost is not determined by output duration alone.
The public weights are distributed under the MiniMax H3 Community License Agreement. The applicable region is worldwide, excluding the European Union, the United Kingdom, South Korea, and the United States—Japan is not among the excluded regions. Organizations wishing to deploy the public weights in excluded regions can apply to MiniMax for a formal license. This is not an open-source license that can be freely used regardless of region.
Commercial use also comes with conditions. Products or services with annual revenue exceeding $20 million must separately obtain prior written approval. Commercial service interfaces must prominently display "MiniMax H3," and the license and NOTICE must be carried forward upon distribution. Using H3's output to improve other AI models is also not permitted, except for H3 itself or its derivative models.
H3 has brought public weights into the top tier of video editing, a space long dominated by proprietary models. However, what developers can choose is not "owning the top-ranked model outright," but rather modifying the 768p core locally or entrusting 2K output and input organization to the API. Going forward, whether Context-IR, Regenerate-2K, and sparse attention will also be released, and how closely the local version can replicate the editing quality of the hosted 2K version given the same instructions, will determine the practical impact of this release.