MiniMax H3

MiniMax H3 (also written MiniMax-H3) is a multimodal video generation model developed by MiniMax, the Shanghai-based AI company that operates the Hailuo AI video service. Announced on July 31, 2026 and released as an open-weight model on Hugging Face on August 3, 2026, H3 generates video with natively synchronized stereo audio at up to 2K resolution and up to 15 seconds, from a context that can mix text, images, video clips and audio clips.^[1][2][3] MiniMax positions it as a general-purpose "omni-modal" system rather than a task-specific video model: text-to-video, image-to-video, video editing, reference-based generation and audio-driven generation are all expressed as natural-language descriptions over a shared multimodal context instead of separate expert models.^[1][3] At the time of the weights release, Artificial Analysis ranked H3 first on its Video Arena video editing leaderboard, which the trade publication The Decoder reported as the first time an open model topped an AI video ranking.^[11][12]

The release is open weights with significant strings attached. The MiniMax H3 Community License Agreement, dated August 2, 2026, defines its "Applicable Territory" as worldwide excluding the European Union, the United Kingdom, South Korea and the United States, meaning users in those jurisdictions are not licensed to run the weights locally without separately obtained authorization; MiniMax's hosted API remains available globally.^[4][5] Two of the three modules of the full H3 system, the H3-Context-IR preprocessing pipeline and the H3-Regenerate-2K resolution stage, were withheld from the initial release, so local deployments top out at 768p output.^[3]

Key facts

Announcement and release

MiniMax announced H3 on July 31, 2026, framing it as a break from the task-fragmented structure of the generative video field: instead of separate expert models for editing, subject reference and motion reference, H3 was pretrained to treat reference and editing relationships as language, on the argument that "language is the bridge to generalization".^[1] The launch post said MiniMax planned "to open up the model weights in the coming days, subject to applicable laws and regulations", and claimed a per-second price at 2K of less than a third of mainstream models; both the openness commitment and the price positioning were company claims at that point.^[1] On the consumer side, the model replaced Hailuo 2.3 as the current model of the Hailuo AI product, whose earlier Hailuo entries MiniMax's platform documentation now lists as legacy models,^[17] and it launched the same day inside third-party creative tools: Krea announced H3 in its video tool on July 31, telling users the model "matches seedance's features at a fraction of the cost", a marketing comparison with ByteDance's Seedance that Krea did not support with published measurements.^[15][16]

The weights followed three days later. On August 3, 2026, MiniMax published the model repository publicly on Hugging Face and ModelScope (the repository had been created privately on July 28) and announced the release under the title "Open General Intelligence: MiniMax H3 Is Now Open Source".^[2][6][8] The Hailuo AI account described it as "an open-weight video model, built for the community and verified to run on GPUs including the RTX 5090 and RTX 6000", a hardware compatibility claim MiniMax made without publishing supporting benchmarks.^[7] The release landed in a crowded news window: The Decoder noted that ByteDance shipped its competing Seedance 2.5 the same day, with support for longer 30-second clips.^[12]

Capabilities

H3 accepts a multimodal context and generates video and audio jointly. The documented output envelope is 4 to 15 second clips at 24 frames per second, in aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, with 32 kHz stereo audio in which speech, sound effects and music are modeled together rather than as separate tracks.^[1][3] The model card states stable dialogue support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish.^[3]

The open release ships as two task-specific checkpoints of the base model, both CFG-distilled and in BF16 precision:^[3]

• H3-Base-FL2VA covers text-to-video and first/last-frame generation: no input image gives text-to-video, one image anchors the first or last frame, and two images anchor both.
• H3-Base-Ref2VA covers omni-reference generation: up to nine reference images, up to three reference video clips (2 to 15 seconds each, 15 seconds total) and up to three reference audio clips, capped at 12 files per request. The launch version of the model card also said reference audio had to be accompanied by an image or video input and could not be the sole input; MiniMax removed that sentence in an August 11, 2026 revision of the card.^[3][33] Reference video plus text instructions also covers video-to-video editing, and reference audio supports voice-timbre-driven speech in the output.

Native output resolution is 768 pixels on the short side. 2K output is produced by a separate stage, H3-Regenerate-2K, in which the base model regenerates its own low-resolution result with the original multimodal context still in hand, in place of a conventional super-resolution module; MiniMax argues this recovers detail such as small text that upscalers can only guess at.^[1][3]

Architecture

The complete H3 system consists of three modules, of which only the middle one is open:^[3]

1. H3-Context-IR, a hosted preprocessing and orchestration system that parses free-form multimodal input, performs cross-modal association and temporal reasoning, and serializes the result into a "Context Intermediate Representation" consumed by the base model. MiniMax describes it as critical to output quality, relies on multiple hosted models to run it, and did not include it in the open release; it is exposed as an API, and prompting guides are provided for developers who want to approximate it.
2. H3-Base, the open-weights generator. It encodes text through the H3-Encoder (the full Qwen3-VL-32B vision-language model, with hidden states drawn from the 50th layer), visual input through both the encoder and a visual VAE, and audio through an audio VAE, then packs everything into a single multimodal token sequence processed by the H3-Omni-Transformer, a 33 billion parameter dense single-stream transformer that predicts video and audio latents jointly. Roughly 13 billion of those parameters sit in AdaLN modulation branches whose outputs can be precomputed and cached, so they need not be loaded for inference-only use. Positional structure uses three-dimensional multimodal rotary embeddings (MM-RoPE) over time, height and width. The model was trained with native sparse attention for long sequences, but the initial release supports full-attention inference only, with the sparse implementation promised later. As of October 30, 2026, the model card still described MiniMax's own sparse-attention implementation as a future update; in the meantime, academic and community projects built their own sparse and hybrid attention schemes for H3 (see the community tooling section below).^[3]
3. H3-Regenerate-2K, the in-context regeneration stage for 2K output, likewise not yet open-sourced and available only through MiniMax's API.

The H3-VisualVAE is a temporally causal video autoencoder with 16x spatial and 4x temporal compression and 24 latent channels; a further 1x2x2 patchification takes the effective spatial downsampling to 32x before the transformer. The H3-AudioVAE compresses each of the two stereo channels of 32 kHz audio into latent tokens at 40 Hz, processing the channels independently through shared encoder and decoder weights.^[3] In the July 31 announcement MiniMax additionally described a captioning approach ("Contextual Omni Representation") that distills roughly 100,000 tokens of inference about source material into about 4,000 tokens of description, credited its rebuilt VAE with a fourfold gain in effective sequence length, and said separating understanding and generation workloads lifted end-to-end training throughput by nearly 30 percent; these are the company's own engineering claims, made ahead of any full technical report.^[1]

License

H3 is released under the MiniMax H3 Community License Agreement, a custom license dated August 2, 2026, issued by the MiniMax entity Nanonoble Pte. Ltd. and governed by Hong Kong law.^[4] Its most unusual term is territorial: the license applies worldwide except in the "Excluded Territories" of the European Union, the United Kingdom, South Korea and the United States, and it states that use, reproduction, modification, distribution or display of the model or its outputs outside the licensed territory is not authorized.^[4] Organizations in excluded regions can apply to MiniMax for a separate formal license, which the company says it grants after reviewing the deployment's compliance controls.^[5]

Other notable terms:^[4]

• Commercial use is permitted, but licensees whose commercial products and services exceed 20 million US dollars in yearly revenue must obtain separate prior written authorization from MiniMax.
• Commercial products or services using H3 must prominently display "MiniMax H3" on their user interface.
• The model and its outputs may not be used to improve any other AI model, other than H3 and its derivatives.
• Redistribution requires passing on the license and a copyright NOTICE file; a "Powered by MiniMax H3" notice and AI-generation identifiers on output files are encouraged but not required.
• MiniMax claims no rights over generated outputs.
• An acceptable use policy bans, among other things, military use, high-risk automated decision-making, impersonation without consent, and undisclosed machine-generated content in public channels.

In an accompanying Q&A, MiniMax explained the territorial carve-out as a response to regulatory exposure specific to video generation, citing the EU AI Act, comparable uncertainty in the UK and South Korea, and, for the US, the fact that the company "is also involved in ongoing copyright-related legal proceedings specifically concerning generative video AI", a reference to the copyright suit filed against MiniMax by Disney, NBCUniversal and Warner Bros. Discovery in September 2025 (covered in the MiniMax article). The company characterized the limitation as meaning "not yet", not "not ever": the hosted API stays globally available, and the license scope is to be revisited as regulation settles.^[5]

A paid licensing channel followed. On August 27, 2026, Comfy Org, the company behind ComfyUI, announced that it had become an official reseller of MiniMax H3 commercial licenses for running the model locally, with a Professional tier listed from $5,000 per month (up to 10 users, distilled open-weight versions) and an Enterprise tier with custom pricing that includes undistilled weights; Comfy said generations on its hosted Comfy Cloud already carry commercial rights.^[30][31] Comfy's announcement and ComfyUI documentation say local commercial use requires such a license, a broader statement than the community license itself, which grants royalty-free commercial use inside the Applicable Territory below the $20 million revenue threshold; Comfy's own license FAQ acknowledges that threshold.^{[4][14][30][31]} Organizations in the four excluded territories can also apply directly through a license request form on MiniMax's developer platform.^[32]

MiniMax's own release post calls H3 "open source", but the combination of territorial limits, the revenue threshold, use restrictions and withheld system components makes "open weights" the more precise description.^[2][4]

Benchmarks and reception

The most-cited independent yardstick is the Artificial Analysis Video Arena, which ranks models by Elo scores from blind pairwise human votes. As of August 5, 2026, in the default with-audio views, H3 ranked first in video editing (Elo 1,130, ahead of Google's Gemini Omni Flash at 1,122), second in text-to-video (Elo 1,239, behind Gemini Omni Flash at 1,244), and third in image-to-video (Elo 1,187, behind ByteDance's Dreamina Seedance 2.0 720p at 1,196 and Gemini Omni Flash at 1,194), with the small gaps at the top of each board leaving nearby models statistically close.^[9][10][11] Artificial Analysis marks H3 as open weights on those boards, and The Decoder reported on August 3 that this made H3 "the first open model to top an AI video ranking".^[11][12] On the text-to-video board, H3's normalized price of $7.80 per minute of 1080p-equivalent output was listed against $9.07 for Seedance 2.0 720p, $20.16 for Kling 3.0 Pro and $21.00 for SkyReels V4.^[9]

A second, unrelated crowdsourced leaderboard, LMArena's Video Arena, points the same way. LMArena announced on August 3, 2026 that H3 (listed there as minimax-h3) had become the top-ranked open model on both its text-to-video and image-to-video boards, about 280 points ahead of the next best open model and a sharp break from MiniMax's earlier Hailuo entries, which it placed in the high 20s.^[18] As of August 7, 2026, LMArena showed minimax-h3 fourth overall in text-to-video (score 1,455, behind Google's Gemini Omni Flash at 1,513, Dreamina Seedance 2.0 720p at 1,479 and Meta's prerelease Muse Video at 1,458) and second overall in image-to-video (1,476, just behind Dreamina Seedance 2.0 720p at 1,478), in each case the highest-placed open-weight model, with the nearest open competitors (hunyuan-video-1.5 and Kandinsky 5.0) roughly 280 points back and MiniMax's own Hailuo 2.3 at rank 28 and 24 respectively.^[19][20] Both LMArena boards were still based on comparatively small H3 vote counts (about 1,000 to 1,200 votes per board, against tens of thousands for longer-listed models), so its position carried wider confidence intervals than incumbents'.^[19][20]

Uptake of the open release was rapid: within two days of publication the Hugging Face repository had accumulated over 1,900 likes and more than 20 community Spaces built on the model.^[8] Day-zero inference support was coordinated with the main open-source serving stacks: the model card links official deployment recipes for SGLang and vLLM, a dedicated diffusers pipeline (MiniMaxH3ModularPipeline), and native ComfyUI workflows for text-to-video, image-to-video and reference-to-video, which require ComfyUI 0.30.0 or later and cap the native canvas at a 768-pixel short edge (up to 768x1344).^[3][8][14] The Decoder noted the practical limits of the release: because H3-Context-IR and the 2K stage remain closed, self-hosted deployments must build their own context preprocessing and cannot reproduce 2K output locally, though the open weights do allow fine-tuning on custom footage, characters or styles.^[12]

Physical-consistency evaluation

The October 6, 2026 preprint World Models' Last Exam in Physics tested eight image-to-video models on 40 controlled tasks, with four samples per task-model pair (1,280 videos). H3 was evaluated at 1344 x 768 pixels with 124 frames and ranked second by average composite score.^[21]

The composite score weights task consistency at 15% and measured physical agreement at 85%; the physical score contributes only when consistency reaches 80/100. The tests assess observable relationships under specified conditions, not general physics knowledge. H3 led the Easy group, while Seedance led the Hard group. These difficulty groups were defined from mean scores across the tested models.^[21]

Availability and pricing

The full three-module system runs on MiniMax's own surfaces: the Hailuo AI web and mobile apps for consumers, and the MiniMax Open Platform API for developers, where H3 is invoked as model MiniMax-H3 through the v2 video generation endpoint, with companion endpoints exposing H3-Context-IR and H3-Regenerate-2K so self-hosters can mix local base-model inference with the hosted stages.^[3] As of August 5, 2026, MiniMax's pay-as-you-go pricing listed H3 output at $0.13 per second at 2K and $0.08 per second at 768p, with input audio free, the first five input images free ($0.04 per additional image), and input video billed by duration at the output resolution's per-second rate.^[13]

Third-party availability began before the weights shipped: Krea offered H3 in its video tool from July 31, 2026, listing 2K output, clip lengths selectable from 5 to 15 seconds and tagged multi-reference input.^[15][16] For self-hosting, the weights are on Hugging Face and ModelScope in both the original checkpoint layout and diffusers format, in BF16, with MiniMax recommending SGLang, vLLM, diffusers or ComfyUI for serving; MiniMax's example SGLang configurations use four GPUs, while its RTX 5090 and RTX 6000 compatibility statement addresses the consumer and workstation end.^[2][3][7]

H3 sits in the same competitive field as ByteDance's Seedance line, Kuaishou's Kling, Google's Veo 3 and successors, OpenAI's Sora, and the open-weight Chinese lines Wan (Wan 2.5 and successors from Alibaba) and Tencent's HunyuanVideo; within the open-weight group, H3 is the first to lead an Artificial Analysis video board.^[9][12]

Community tooling and acceleration

The open release drew a large ecosystem of derivative checkpoints, quantizations and inference tools on Hugging Face. Most of them target the main practical problem with running H3 locally: the base model calls its 33 billion parameter transformer dozens of times per clip over long audio-video token sequences, so generation on a single GPU is slow.^[23] Two techniques dominate: step distillation, which cuts the number of transformer calls, and sparse attention, which cuts the cost of each call. Companies have also released post-trained derivatives: on October 29, 2026 Pika published Pika H3 Flex, two LoRA adapters that generate in six sampling steps and that its model card describes as "Model Derivatives of MiniMax H3", distributed under the MiniMax H3 Community License Agreement.^[36][37]

ComfyUI support

ComfyUI shipped native H3 support in its v0.30.0 release on August 3, 2026, the day the weights went public, adding both local model support (contributed by the developer kijai) and hosted API partner nodes; Comfy Org also publishes repackaged H3 weights and official workflow templates for text-to-video, image-to-video and reference-to-video.^[14][24] Later releases added multiframe reference anchoring, a Fun ControlNet Union workflow and, from ComfyUI 0.36.0, templates for FastVideo's distilled FastH3 checkpoint.^[14][25] A core Block Sparse Attention node, merged on September 6, 2026, lets ComfyUI workflows run H3 with sparse attention, including a Video Sparse Attention mode that reads trained gate layers from FastH3 weights and a Sol-Attn mode that works on any H3 checkpoint.^[25][26]

Video Sparse Attention and FastH3

Video Sparse Attention (VSA) comes from the paper "VSA: Faster Video Diffusion with Trainable Sparse Attention" by Peiyuan Zhang and colleagues, including Ion Stoica, Eric Xing and Hao Zhang (arXiv:2505.13389, May 2025, accepted at NeurIPS 2025).^[22] VSA splits a video's tokens into cubes of 4x4x4 tokens. A coarse stage pools each cube into a single representation and computes attention between cubes to find the most important ones; a fine stage then computes ordinary token-level attention only inside the selected cubes, and a learned gate combines the two. Because the whole mechanism is trainable, it can replace full attention during training as well as inference; retrofitting it into Wan 2.1 cut that model's end-to-end generation time from 31 to 18 seconds in the authors' tests.^[22]

The FastVideo project at Hao AI Lab (UC San Diego), whose repository hosts the VSA code, applied the technique to H3. On August 27, 2026 it released FastH3 Preview v1, a text-to-video-and-audio checkpoint post-trained with DMD2 distillation to need four transformer calls instead of the base model's 49, with VSA keeping about 10% of video-to-video tiles while text and audio stay dense.^[23] On one NVIDIA B200, FastVideo reported 16.2 seconds end to end for a 5-second 1344x768 clip against 132.5 seconds for dense base H3 (8.16x), and up to 14.38x for 15-second clips; these are the team's own measurements.^[23] That preview did not cover first/last-frame or reference-to-video generation, because Ref2VA uses a separate reference transformer.^[23] FastVideo released FastH3 Omni, an eight-step distilled reference-to-video model, on October 20, 2026.^[34]

Video DeltaNet

Video DeltaNet (VDN), described in a September 17, 2026 arXiv paper by Haocheng Xi, Kurt Keutzer, Xiuyu Li and colleagues, takes a different route: it pairs local softmax attention with a bidirectional linear attention memory that is updated once per frame, applying the hybrid only to video-to-video interactions and keeping softmax attention for text and audio.^[29] Instantiated on H3 as VDN-H3 with eight-step distillation and an optimized SGLang serving stack, it finished denoising a 14.3-second 768p video in 6.70 seconds on eight B200 GPUs, which the authors report as a 14.5x speedup over the 50-step dense baseline on the same hardware.^[29]

ComfyUI-Ref2VA-VSA

ComfyUI-Ref2VA-VSA is a community custom node, published on GitHub by the user Kablex on September 5, 2026, that applies VSA to H3's reference-to-video (Ref2VA) path, which FastH3 Preview v1 did not cover.^[27] Ref2VA places text, reference-image and reference-audio tokens ahead of the generated video sequence, and the README says naive 3D tiling mixes tokens from different modalities in one tile. The node instead keeps those conditioning tokens in separate tiles that are never pruned, applies top-k sparsity only to generated-video tiles, and transplants the 50 trained VSA gate projections from a FastH3 checkpoint onto the Ref2VA model.^[27][28] It is paired with a four-step Turbo LoRA from the LightX2V project.^[27][35]

The author reports the following comparison (the README's two showcased videos share one seed, prompt and reference image) on one GeForce RTX 4090 (24 GB) for a 1344x768, 124-frame, 5.16-second clip at 24 fps. These are the repository author's own figures from a single prompt and reference image, not independent benchmarks:^[27]

The README's headline "9x" figure compares the four-step sparse run with the 50-step dense baseline, so by its own table most of that gain comes from step distillation; against the dense four-step Turbo LoRA run listed in the same table (about 190 s), the sparse run (about 72 s) is roughly 2.6 times faster by this article's arithmetic, though the README does not say whether the two runs used the same checkpoint and quantization. The README lists a 24 GB NVIDIA GPU as the requirement.^[27] As of October 30, 2026, the repository had 878 stars and 103 forks on GitHub.^[27]

Continue reading

• MiniMaxMiniMax (稀宇科技; listed as MiniMax Group Inc.) is a publicly traded artificial intelligence (AI) company headquartered in Shanghai, China that develops large language models, multimodal AI systems, and…Mentioned 53 times above
• ComfyUIComfyUI is a free, open-source node-based graphical user interface and backend for generative AI workflows, primarily focused on image generation using diffusion models such as Stable Diffusion and FLUX.Mentioned 11 times above
• OOpen weightsOpen weights is the practice of publishing the trained parameter tensors of a neural network so that anyone can download, run, fine-tune, and redistribute the model without obtaining a separate inference…Mentioned 11 times above
• SeedanceSeedance is the family of foundation video generation models built by the Seed team at ByteDance, the Chinese internet company that owns TikTok and Douyin.Mentioned 10 times above
• SSparse attentionSparse attention is a family of techniques that cut the computational and memory cost of the attention mechanism in transformer models by letting each token attend to only a subset of other tokens in a sequenceMentioned 8 times above
• Hailuo AIHailuo AI (海螺AI) is the consumer AI video generation service operated by MiniMax, a Shanghai-based foundation model developer that has traded on the Hong Kong Stock Exchange as MiniMax Group Inc. (stock code…Mentioned 4 times above

More from See also

• H3 Max
• Artificial Analysis
• AI Video Generation

More to read

• Pika Labs
• Pika H3 Flex
• Seedance 2.0

Topics:AI ModelsChinese AIOpen Source AIVideo Generation

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8MiniMax. "MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities." July 31, 2026. minimax.io/...minimax-h3
2. ^1 ^2 ^3 ^4 ^5MiniMax. "Open General Intelligence: MiniMax H3 Is Now Open Source." August 3, 2026. minimax.io/...minimax-h3-open-source
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18MiniMaxAI. "MiniMax-H3" model card. Hugging Face. Accessed August 5, 2026. huggingface.co/...MiniMax-H3
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9MiniMax (Nanonoble Pte. Ltd.). "MiniMax H3 Community License Agreement." August 2, 2026. huggingface.co/...LICENSE
5. ^1 ^2 ^3MiniMax. "Q&A About License." MiniMax-H3 repository. Accessed August 5, 2026. huggingface.co/...QA-about-License.md
6. ^1 ^2MiniMax (@MiniMax_AI). "MiniMax-H3 Is Now Publicly Available." X, August 3, 2026. x.com/...2084106804032872591
7. ^1 ^2Hailuo AI (@Hailuo_AI). "An open-weight video model, built for the community and verified to run on GPUs including the RTX 5090 and RTX 6000." X, August 3, 2026. x.com/...2084123156818915727
8. ^1 ^2 ^3 ^4 ^5Hugging Face. API metadata for MiniMaxAI/MiniMax-H3. Accessed August 5, 2026. huggingface.co/...MiniMax-H3
9. ^1 ^2 ^3Artificial Analysis. "Text to Video Leaderboard." Accessed August 5, 2026. artificialanalysis.ai/...text-to-video
10. ^Artificial Analysis. "Image to Video Leaderboard." Accessed August 5, 2026. artificialanalysis.ai/...image-to-video
11. ^1 ^2 ^3Artificial Analysis. "Video Editing Leaderboard." Accessed August 5, 2026. artificialanalysis.ai/...video-editing
12. ^1 ^2 ^3 ^4 ^5Schreiner, Maximilian. "China's MiniMax H3 is the first open model to top an AI video ranking." The Decoder, August 3, 2026. the-decoder.com/...odel-to-top-an-ai-video-ranking
13. ^1 ^2MiniMax API Docs. "Pay as You Go." Accessed August 5, 2026. platform.minimax.io/...pricing-paygo
14. ^1 ^2 ^3 ^4Comfy Org. "ComfyUI MiniMax H3 Video Generation Guide" (originally "ComfyUI MiniMax H3 Native Workflow Tutorial"). ComfyUI documentation. Accessed October 30, 2026. docs.comfy.org/...minimax-h3
15. ^1 ^2Krea (@krea_ai). "introducing MiniMax H3." X, July 31, 2026. x.com/...2083014535481598410
16. ^1 ^2Krea. "MiniMax H3 by MiniMax, AI Video Generator." Accessed August 5, 2026. krea.ai/...minimax-h3
17. ^MiniMax API Docs. "Models." Accessed August 5, 2026. platform.minimax.io/...models-intro
18. ^Arena.ai (@arena). "Big news: MiniMax-H3 by @MiniMax_AI is now the #1 open model in Video Arena: across both Text-to-Video and Image-to-Video." X, August 3, 2026. x.com/...2084408459991421319
19. ^1 ^2LMArena. "Text to Video Leaderboard." Accessed August 7, 2026. lmarena.ai/...text-to-video
20. ^1 ^2LMArena. "Image to Video Leaderboard." Accessed August 7, 2026. lmarena.ai/...image-to-video
21. ^1 ^2 ^3 ^4Gao, Mingju, et al. "World Models' Last Exam in Physics." arXiv:2610.08791v1, October 6, 2026. arxiv.org/...2610.08791v1
22. ^1 ^2Zhang, Peiyuan; Chen, Yongqi; Huang, Haofeng; Lin, Will; Liu, Zhengzhong; Stoica, Ion; Xing, Eric; Zhang, Hao. "VSA: Faster Video Diffusion with Trainable Sparse Attention." arXiv:2505.13389, May 19, 2025 (NeurIPS 2025). arxiv.org/...2505.13389
23. ^1 ^2 ^3 ^4FastVideo Team, Hao AI Lab @ UCSD. "FastVideo FastH3 V1: Open-Weight 4-Step Sparse Distilled Minimax H3 for 14x Speedup on NVIDIA Blackwell GPU." August 27, 2026. haoailab.com/...fasth3-preview
24. ^Comfy-Org. "ComfyUI v0.30.0" release notes. GitHub, August 3, 2026. github.com/...v0.30.0
25. ^1 ^2Comfy Org. "FastVideo FastH3: ComfyUI Workflow Examples." ComfyUI documentation. Accessed October 30, 2026. docs.comfy.org/...minimax-h3-fastvideo
26. ^kijai. "Add Sparse Attention node." ComfyUI pull request #16072, merged September 6, 2026. github.com/...16072
27. ^1 ^2 ^3 ^4 ^5 ^6Kablex. "ComfyUI-Ref2VA-VSA" (README and repository metadata). GitHub. Accessed October 30, 2026. github.com/...ComfyUI-Ref2VA-VSA
28. ^ComfyUI Wiki. "Ref2VA VSA: Sparse Attention Node for H3 Reference-to-Video." Accessed October 30, 2026. comfyui-wiki.com/...2026-09-19-ref2va-vsa-node
29. ^1 ^2Xi, Haocheng, et al. "Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation." arXiv:2609.20744, September 17, 2026. arxiv.org/...2609.20744
30. ^1 ^2Villegas, Josiah. "Comfy Is Now an Official Reseller of MiniMax H3 Commercial Licenses." Comfy blog, August 27, 2026. blog.comfy.org/...s-now-the-only-official-reseller
31. ^1 ^2Comfy Org. "MiniMax Commercial License." Accessed October 30, 2026. comfy.org/...license
32. ^MiniMax. "MiniMax H3 License Request Form." Accessed October 30, 2026. platform.minimax.io/h3-license
33. ^MiniMaxAI. MiniMax-H3 repository commit history (README revisions of August 10 and August 11, 2026). Hugging Face. Accessed October 30, 2026. huggingface.co/...main
34. ^FastVideo Team, Hao AI Lab @ UCSD. "FastVideo FastH3 Omni: Open Source Real-Time Minimax FastH3 Ref2VA on NVIDIA Blackwell, RTX Spark, and RTX GPUs." October 20, 2026. haoailab.com/...fastvideo-omniref-update
35. ^LightX2V. "MiniMax-H3 Turbo." Hugging Face. Accessed October 30, 2026. huggingface.co/...Minimax-h3-Turbo
36. ^Pika. "Pika-Labs/Pika-H3-Flex." Hugging Face model card. Accessed November 1, 2026. huggingface.co/...Pika-H3-Flex
37. ^Pika Team. "Introducing Pika H3 Flex." Pika blog, October 29, 2026. pika.art/...pika-h3-flex

Show all 37 references