What Is MiniMax H3
MiniMax H3 was released on July 31, 2026 under the name Hailuo 3.0 and is a 33-billion-parameter dense omni-modal transformer. It treats text, images, video, and audio as a single stream rather than stitching together multiple independent pipelines. The model generates 4- to 15-second clips at 24 FPS with 32 kHz stereo, reaching 768 pixels on the short side locally and up to 2K through a hosted regeneration stage. Counting the encoder and VAE, the full inference stack is close to 69 billion parameters.
Key Capabilities
Sound is generated, not added in post-production. The H3-Omni Transformer produces vocals, sound effects, and music in the same generation, officially recorded as 32 kHz stereo, eliminating an entire production stage that models requiring a separate audio step need.
H3-VAE delivers roughly a 4x increase in sequence length, which is the structural reason 15-second clips with synchronized sound are feasible. H3-Contextual Omni Representation uses language as a bridge between modalities, and an average of roughly 100,000 tokens of inference is compressed to about 4K before entering generation.
2K is achieved through regeneration rather than upscaling. The base model renders 768 pixels on the short side, and H3 regenerates within the original context to reach 2K. The hosted API can directly request 2K; local deployment outputs 768p unless you run the regeneration stage yourself.
Specification Highlights
- 33-billion-parameter dense omni-modal transformer, with a full inference stack of about 69 billion parameters including the encoder and VAE
- Integer durations of 4 to 15 seconds, fixed 24 FPS, with native multi-shot modeling
- Native 32 kHz stereo, with vocals, sound effects, and ambient sound generated alongside the visuals
- Aspect ratio examples listed in the official documentation include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16
- Input paths in the documentation include FL2VA (text prompt plus 0, 1, or 2 images) and Ref2VA (mixed image, video, and audio references)
- Ref2VA allows up to 9 images, 3 videos, and 3 audio files, but the combined total file count is limited to 12, and audio must be paired with images or video
- Prompt limit of 7,000 characters, request body limit of 64 MB
Usage Methods
MiniMax H3 has three mainPath:
- Hailuo AI app: the consumer app offers H3 through a credit system, with a free tier and monthly paid plans; the latest rates need to be checked in-app. This path is the fastest way to see the model's performance, but it does not expose endpoint, random seed, or resolution parameters.
- Open platform API: billed by output second, the 2K tier is higher than the 768p tier, and the model ID is MiniMax-H3. The Context-IR variant is priced separately by token. The 768p tier was once intermittently restricted, so the real-time price list should be confirmed before bulk budgeting.
- Local weights: checkpoints released in early August 2026. The minimum usable set is about 42.5 GB, the two checkpoints at minimum usable precision are about 63.4 GB, and the full bf16 research weights are about 123.6 GB. ComfyUI reports that under aggressive offloading, 12 GB of VRAM is the lower bound, and 64 GB of system memory makes running more acceptable.
Applicable scenarios
According to official materials, MiniMax H3 is suitable for short films, product showcases, character actions, vertical short dramas, fashion imagery, and other scenarios that require image and sound generation in one pass. The official documentation provides nine reproducible prompt examples, covering uses such as sci-fi trailers, animated food vignettes, vertical short dramas, product launches, character actions, reference-led scene editing, vertical anime characters, fashion imagery, and dialogue replacement.
Restrictions
MiniMax H3's community license excludes the United States, the European Union, the United Kingdom, and South Korea, and hosted access was also not launched in these markets at release time; MiniMax mentions evolving AI regulation and copyright concerns. Within these regions, MiniMax H3 is better understood as a research object rather than a deployment object, and licensed alternatives such as Sora 2, Veo 3.1, or Seedance 1.5 Pro may be considered.
The weights are downloadable, but they are not open source in the OSI sense. Commercial use requires a visibly displayed "MiniMax H3" attribution in the interface, model distillation is prohibited, and organizations with annual revenue of about 20 million USD or more require separate written authorization.
Other restrictions: local deployment defaults to 768p output unless the regeneration stage is run separately; the consumer app does not expose endpoint, seed, or resolution parameters, which is unfavorable for reproducible comparisons; audio reference needs to be paired with image or video; when building reproducible evaluations, the provider, endpoint, variant, and resolution need to be recorded.
Pricing
MiniMax's open platform bills by output seconds, the 2K tier is higher than the 768p tier, and the Context-IR variant is billed by token, and the specific amounts need to be checked in the real-time price list. The pricing page on the site lists credit packages: the Basic annual plan is 800 credits per month and 9,600 credits for the full year; the Pro annual plan is 2,000 credits per month and 24,000 credits for the full year; the Enterprise annual plan is 6,000 credits per month and 72,000 credits for the full year, and it also lists the corresponding monthly image limits. The materials note that on-site generation is metered in credits: 18 credits per second for 768p and 29 credits per second for 2K.
