MiniMax H3 launched in July 2026 as a 50-layer omni-modal transformer, then open-sourced its weights on Hugging Face in early August 2026. It replaces MiniMax's prior Hailuo video model for teams that need synchronized audio-video output or want to self-host outside the US, EU, UK and South Korea.
MiniMax H3 is an open-weight, omni-modal video generation model from MiniMax that generates 2K video clips with synchronized stereo audio produced in the same forward pass, a capability rivals such as Veo or Kling do not offer natively since they generate silent video without joint audio synthesis.
Provider: MiniMax · Family: Hailuo 3
Input modalities: text, image, video, audio · Output: video, audio
About MiniMax H3
MiniMax H3 is an omni-modal video generation model built by MiniMax, the Shanghai AI lab, and released on July 31, 2026 through the MiniMax platform API and the consumer Hailuo AI app. Branded to consumers as Hailuo 3.0, it is the direct successor to Hailuo 2.3 and the first model in the lineup to treat audio as a first-class generated modality rather than a bolted-on post-process. Where the previous generation topped out at 1080p and roughly 10 seconds of silent or separately-dubbed video, H3 produces native 2K output at 24fps for clips up to 15 seconds, extendable to around 30 seconds through a dedicated regeneration pass. Architecturally, MiniMax discarded the prior design entirely rather than iterate on it, on the stated reasoning that task-specific tricks for video generation do not generalize across modalities. The result is H3-Omni-Transformer: a dense, single-stream, 50-layer transformer with a hidden size of 5,376 and 56 attention heads, using three-dimensional rotary position embeddings across time, height and width so text, image, video and audio tokens share one sequence and one set of weights. The public system is actually three cooperating checkpoints: H3-Context-IR turns free-form multimodal prompts into a structured intermediate representation, H3-Base generates a synchronized audio-video draft at lower resolution, and H3-Regenerate-2K upsamples that draft to final 2K using the original context plus the low-res result. H3 has no context window in the token-based LLM sense; its multimodal context instead comes from omni-reference inputs, up to 9 reference images, 3 video clips and 3 audio clips combined in a single generation call to hold character identity, art style, camera motion and voice consistent across shots. It also restores first-and-last-frame control (one image drives first-frame-to-video or last-frame-to-video; two images drive full first-and-last-frame interpolation), a capability the prior Hailuo generation had dropped, rebuilt here on a cleaner input model. Independent VBench-style leaderboard scores were not published at launch, so quality comparisons against Google Veo, OpenAI Sora 2 and Kling remain qualitative rather than benchmarked: reviewers consistently rate Veo ahead on pure photorealism and Sora 2 ahead on temporal stability, while crediting H3 as the only one of the three with genuine joint audio-video generation and reasoning over audio inputs. H3 is billed by seconds of rendered output rather than by tokens, with separate rates for its two output tiers (see pricing below). It is reachable through the official MiniMax platform API, the Hailuo AI consumer app, and third-party inference hosts including OpenRouter, Segmind, fal, Atlas Cloud and EvoLink, several of which added support within a day of launch. On August 3, 2026, MiniMax additionally open-sourced the weights (MiniMaxAI/MiniMax-H3 on Hugging Face, Diffusers format, roughly 42.5GB) under the MiniMax H3 Community License Agreement, covering two task-specific checkpoints: FL2VA for text-to-video and frame-conditioned generation, and Ref2VA for reference-conditioned generation from images, video and audio. Community quantizations followed almost immediately, ranging from a compact NVFP4 build up to a full-precision INT8 build (see hardware requirements below), plus native ComfyUI packaging. The license carve-out is the sharpest edge of the release: the MiniMax H3 Community License Agreement, effective August 2, 2026, explicitly excludes the United States, the European Union, the United Kingdom and South Korea from its definition of Applicable Territory, meaning users physically in those regions are not licensed to run, modify or distribute the open weights locally, citing evolving AI regulation and copyright exposure for generative video. The hosted API remains globally available with built-in safeguards regardless of this restriction, so teams in restricted regions can still use H3 through MiniMax's own endpoint or a third-party API host, just not as a self-hosted deployment. MiniMax has not published a system card, training-data breakdown, or content-moderation disclosure for H3 as of this writing; training data composition and cutoff are undisclosed, consistent with the rest of the MiniMax model family. Buyers who need documented red-teaming, a published refusal-rate benchmark, or SOC2/HIPAA-style compliance attestations will not find them here yet. H3 fits teams building short-form video ads, social content, dubbing and localization workflows (thanks to native multi-language stereo audio), and local ComfyUI-based pipelines who want an open-weight alternative to closed video APIs. It is the wrong pick for anyone needing benchmarked output quality guarantees, EU/UK/US/Korea-based self-hosting under the current license, or any text/code/reasoning workload, since H3 is a video generation model only and has no chat, tool-use, or LLM-style benchmark scores.
Pricing
Billed per second of rendered video, not per token. MiniMax's own pay-as-you-go tier: $0.13/sec at 2K, $0.09/sec at 768p. Third-party hosts price close to that (OpenRouter from $0.13/sec at 2K; Segmind around $0.16/sec at 2K). A short 2K clip typically costs well under a dollar, and even the model's longest supported clip stays under $2.50 depending on host.
Key Features
- Native Stereo Audio: Generates dialogue, sound effects, and room tone in the same forward pass as the video, no separate TTS or dubbing stage required.
- Native 2K Output: Renders extended-length clips at 2K resolution and a cinema-standard frame rate, well past the previous Hailuo generation's 1080p ceiling.
- Omni-Reference Control: Combines several reference images with short reference video and audio clips in a single call, keeping character identity, style, motion and voice consistent across shots.
- First-and-Last-Frame Control: A single reference frame can seed the opening or closing shot; supplying both endpoints generates a full interpolation between them, a control mode the previous Hailuo generation lacked.
- Open Weights with Local Quantization: Open-weight checkpoints available on Hugging Face with community NVFP4, mixed INT4/INT8 and INT8 quantizations for local GPU deployment.
Pros
- Genuine joint audio-video generation, not a separate dubbing pass, a real differentiator versus Veo and Kling.
- Open-weight release with quantizations that fit on a single consumer GPU, unusual for a frontier-tier video model.
- Restored first-and-last-frame control plus 9-image/3-video/3-audio omni-reference gives more compositional control than most 2026 video APIs.
Cons
- Open-weight local deployment is legally off-limits in the US, EU, UK and South Korea under the current Community License.
- No independently verified quality benchmarks (VBench or otherwise) published at launch.
- No system card or content-moderation disclosure, so safety posture cannot be independently assessed.
Frequently Asked Questions
How much does MiniMax H3 cost?
MiniMax H3 bills per second of rendered video rather than per token: $0.13 per second at 2K and $0.09 per second at 768p on MiniMax's own pay-as-you-go tier. Third-party hosts like OpenRouter and Segmind price within a few cents of that rate. A full 15-second 2K clip costs roughly $2.00, while the open-weight release can be self-hosted for the cost of local GPU compute where the license permits it.
How does MiniMax H3 compare to Veo, Sora, and Kling?
Reviewers generally rate Google Veo 3.1 ahead on pure cinematic photorealism and OpenAI Sora 2 ahead on temporal stability, while Kling is known for strong physics and character consistency but produces silent video. MiniMax H3's distinguishing edge is native, synchronized stereo audio generated in the same pass as the video, which none of those three do natively. No independently published VBench scores existed for H3 at launch, so these are qualitative comparisons, not benchmarked rankings.
Is MiniMax H3 open source or proprietary?
MiniMax H3 is open-weight, not fully open-source. MiniMax released the 33B-parameter weights (about 42.5GB) on Hugging Face on August 3 2026 under the MiniMax H3 Community License Agreement, which explicitly excludes the US, EU, UK and South Korea from local deployment rights. The hosted API remains available globally regardless of the local-deployment restriction.
Does MiniMax H3 train on user data?
MiniMax has not published a data-retention or training-on-inputs policy for H3 as of August 2026. No system card or training-data breakdown has been released, so this cannot be independently verified; teams with strict data-handling requirements should confirm directly with MiniMax before sending sensitive reference material.
Who is MiniMax H3 best for and who should avoid it?
H3 suits teams making short-form video ads, social content, or localized/dubbed video who want synchronized audio generated automatically, and ComfyUI users who want an open-weight video model to self-host. Teams based in the US, EU, UK or South Korea wanting to self-host should avoid it under the current license and use the API instead, and anyone needing independently benchmarked output quality or documented safety posture should wait for further disclosures.