MiniMax H3 is available now through MiniMax's official global API and as downloadable base-model weights. The official API generates 4–15 seconds of 768P or 2K video with native stereo audio. MiniMax's Hugging Face repository now publishes separate FL2VA and Ref2VA checkpoints for local 768P generation under the MiniMax H3 Community License.
The complete hosted system is not fully local. H3-Context-IR and H3-Regenerate-2K remain API services, and the first public local release does not include MiniMax's sparse-attention implementation. This page separates those states so “open weights” is not mistaken for a fully open, offline 2K pipeline.
H3, H3 Max, or H3 Max Turbo?
| Choose | When it fits | Access and limit |
|---|---|---|
| MiniMax H3 | You need 2K delivery, hosted editing, or a local 768P base model | MiniMax API or fal; FL2VA/Ref2VA base weights are released, but the full 2K pipeline remains hosted |
| fal H3 Max | You need fast previews or multimodal reference-to-video | fal hosted; 480p, 768p, or 1080p; no H3 Max editing endpoint documented |
| fal H3 Max Turbo | You favor faster, lower-cost text or image drafts | fal hosted; 480p, 768p, or 1080p; no dedicated multimodal reference endpoint documented |
H3 Max and Turbo are separate fal variants, not downloadable versions of MiniMax's H3-Base weights. Check the selected endpoint's live price and limits before committing a production budget.
MiniMax H3 Quick Facts
| Item | Verified status on September 19, 2026 |
|---|---|
| Release announcement | July 31, 2026 |
| Official model ID | MiniMax-H3 |
| Inputs | Text, images, video, and audio in one reference context |
| Output | 768P or 2K video at 24 FPS with 32 kHz stereo audio |
| Duration | 4–15 seconds through the official MiniMax API |
| Reference limits | Up to 9 images, 3 videos, and 3 audio clips; 12 files total |
| Generation modes | Text-to-video, first/last-frame, multimodal reference-to-video, and editing |
| Official API | Available globally at POST /v2/video_generation |
| Official output price | $0.08/sec at 768P; $0.13/sec at 2K |
| Open weights | FL2VA and Ref2VA H3-Base checkpoints published on Hugging Face |
| Local output | 768P H3-Base; the official 2K regeneration module is not yet downloadable |
What Is MiniMax H3?
MiniMax H3 is a general-purpose audiovisual generation model. Instead of treating text-to-video, motion transfer, voice reference, and local editing as unrelated tools, it reads multiple media types as a shared context and returns a synchronized video and audio result.
That design matters when a shot needs more than a prompt. A creator can use one image for character identity, another for art direction, a video for movement, an audio clip for voice or rhythm, and text for camera and editing instructions. H3 is designed to resolve those references together.
MiniMax positions the model for advertising, branding, ecommerce, product design, UI motion, games, and other commercial creative work. Those use cases match its emphasis on readable text, product preservation, motion transfer, and targeted editing. Narrative production is still possible, but teams should test character continuity and causal action across several shots rather than infer episode-level reliability from launch demos.
MiniMax H3 Specs and Reference Limits
MiniMax's official V2 API exposes three H3 input patterns through one endpoint:
- Text to video: a prompt generates a 4-to-15-second 768P or 2K clip.
- Image to video: one image becomes the opening frame, with an optional final frame for transition control.
- Reference to video: up to 9 images, 3 video clips, and 3 audio tracks can guide identity, style, motion, camera, or sound.
The reference request caps the complete upload set at 12 files. Reference video and audio are each limited to 15 seconds in total. The official API accepts adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 output ratios where applicable; first/last-frame generation follows the input image.
Hosted providers can expose different schemas, duration ranges, or prices, so record the provider and exact endpoint with every output.
Native Audio and Editing
Every documented H3 generation includes stereo audio. The launch material describes dialogue, music, ambience, and sound effects timed to the picture, plus voice transfer from an audio reference.
H3 also promotes instruction-based editing: replacing a product, rewriting signage, changing dialogue, relighting a scene, or adding and removing objects while preserving untargeted content. This is more production-relevant than regenerating an otherwise approved clip to fix one detail.
Treat both capabilities as testable claims. A useful acceptance check measures lip synchronization, voice identity, unwanted music, text stability, whether the edit leaks outside the selected subject, and whether a second edit preserves the first.
MiniMax H3 Demo and Hands-On Video
MiniMax and fal also publish selected examples on the MiniMax H3 launch page. Those examples show typography, interface motion, character work, motion reference, compositing, and video editing. Because they are launch selections, use them to understand supported workflows—not to estimate first-try success rate.
MiniMax H3 Official API Access
MiniMax's global API accepts asynchronous H3 jobs at POST https://api.minimax.io/v2/video_generation. Use the official model ID MiniMax-H3, send the prompt and media references in the content array, and query the returned task_id for status and output. The API currently requires a pay-as-you-go key rather than a Token Plan key.
fal also exposes minimax/h3/text-to-video, minimax/h3/image-to-video, and minimax/h3/reference-to-video. Those are provider paths, not MiniMax's official endpoint names. Store the provider, model ID, request ID, full input manifest, price at submission, and output URL with each production job.
MiniMax H3 API Pricing
MiniMax's official pay-as-you-go list price on September 19 is $0.08 per output second at 768P and $0.13 per output second at 2K:
| Duration | 768P output | 2K output |
|---|---|---|
| 5 seconds | $0.40 | $0.65 |
| 10 seconds | $0.80 | $1.30 |
| 15 seconds | $1.20 | $1.95 |
Audio references and the first five images are free. Each additional image costs $0.04. Reference video is billed per input second at the selected output rate. Regenerating a 768P result to 2K costs $0.05 per output second, and the original input materials are billed again under the regeneration rules.
fal's public H3 page currently shows the same $0.08/second 768P and $0.13/second 2K output rates, plus a $0.16/second 4K option that is not listed in MiniMax's official H3 API. Treat every price as a dated provider snapshot. Production budgets should include reference-video charges, retries, storage, review, and editing, then compare cost per approved second.
Are MiniMax H3 Weights Available?
Yes. The official MiniMaxAI/MiniMax-H3 Hugging Face repository now publishes BF16 H3-Base checkpoints for two task families:
- FL2VA: text-to-audio-video plus optional first frame, last frame, or both;
- Ref2VA: text with reference images, videos, and/or audio.
The repository includes the processor, tokenizer, text encoder, visual VAE, audio VAE, and Omni Transformer. Primary runtime guides now cover SGLang, vLLM-Omni, diffusers, and ComfyUI. Their published hardware profiles are examples rather than a universal minimum, so size a deployment from the chosen runtime, checkpoint, resolution, and measured memory use.
MiniMax's official ModelScope account now mirrors the H3 model card and downloadable release. Treat Hugging Face and ModelScope as two distribution paths for the same official model, not as separate H3 variants.
What Is Not Yet Fully Open or Offline?
The official model card identifies three important boundaries:
- H3-Context-IR is hosted. The multimodal instruction-refinement system is not included in the weight release; MiniMax provides an API and prompting guidance instead.
- H3-Regenerate-2K is hosted. Local H3-Base output is 768P. Reproducing the official 2K workflow currently calls MiniMax's regeneration API.
- Sparse attention is pending. The first release uses full-attention inference; MiniMax says its sparse-attention implementation will be published later.
This is an open-weight release with a documented license and local 768P path, not a completely offline copy of the hosted 2K stack.
Who Should Test MiniMax H3?
H3 belongs on the shortlist when a workflow needs 2K output, native audio, first/last-frame control, motion reference, several reference modalities, targeted editing, or accurate on-screen graphics.
For motion comics and short drama, run a sequence benchmark with at least:
- two dialogue close-ups using the same character and voice;
- a full-body action shot with hand-to-object contact;
- two-character staging with a camera move;
- a vertical shot that preserves costume and background;
- one local edit to a prop or sign;
- one shot that must match an earlier approved frame.
AniKuku's production workspace keeps scripts, shots, character assets, scene references, prompts, and output versions outside the provider. That makes it possible to test H3 on selected shots without rebuilding the episode around one API.
See the AniKuku showcase for the output-review format used across model tests.
Verification Method and Machine-Readable Status
This status page uses a simple evidence order: MiniMax API documentation and the official MiniMax Hugging Face and ModelScope repositories first, then named hosted-provider documentation for provider-specific price and schema differences. A capability is marked local only when downloadable weights, a license, and a documented runtime path are all present. Hosted-only modules remain labeled hosted even when they participate in a hybrid local workflow.
The same facts are available in the versioned MiniMax H3 status JSON. The file includes source URLs, verification dates, API prices, released checkpoints, hosted-only components, and runtime links so researchers and publishers can cite or diff the status without scraping this article.
Update Log
- September 19, 2026: Reconfirmed the official API, prices, public checkpoints, and hosted-only Context-IR/2K regeneration; added primary SGLang, vLLM-Omni, diffusers, ComfyUI, and official integration-catalog sources. Sparse attention remains pending.
- September 11, 2026: Verified MiniMax's official ModelScope mirror and reconfirmed the API, prices, checkpoints, ComfyUI workflow, hosted-only Context-IR/2K regeneration, and pending sparse attention.
- August 12, 2026: Verified the official global H3 API, pay-as-you-go pricing, Hugging Face checkpoints, community license, local 768P runtimes, hosted-only Context-IR/2K regeneration, and official ComfyUI workflow.
- August 1, 2026: Published the initial launch, hosted API, specification, and open-weight announcement review.
FAQ
Is MiniMax H3 available now?
Yes. MiniMax's official global API, official web app, public H3-Base weights, and fal's hosted API were all available when checked on September 19, 2026.
Is MiniMax H3 the same as Hailuo 3.0?
Use MiniMax-H3 for MiniMax's official V2 API and preserve the exact provider model ID elsewhere. Product labels and provider aliases do not guarantee identical defaults, pricing, output formats, or safety layers.
Does MiniMax H3 generate audio?
Yes. The current launch and API materials describe native stereo audio for every generation, including dialogue, music, effects, and ambience.
Can MiniMax H3 run locally?
Yes for H3-Base generation at 768P using the public FL2VA or Ref2VA checkpoints. The official Context-IR and 2K regeneration stages remain hosted services, and the initial public inference path does not include sparse attention. Teams with private-cloud or on-premises requirements can review the AniKuku + MiniMax H3 private deployment guide for architecture options and prerequisites.
Is MiniMax H3 better than Seedance?
Both H3 and Seedance 2.5 now have documented hosted access. MiniMax's official H3 API offers 2K output; the current BytePlus Seedance 2.5 API supports longer clips and more references but documents 480p/720p output. The best choice depends on reference control, continuity, audio, retry rate, latency, and cost. See the Seedance vs MiniMax H3 comparison.
Sources
- MiniMax official H3 announcement
- MiniMax official H3 features thread
- MiniMax H3 official API reference
- MiniMax official pay-as-you-go pricing
- Official MiniMax H3 weights, license, and deployment guide
- Official MiniMax H3 ModelScope mirror
- Official MiniMax H3 ComfyUI workflow
- SGLang MiniMax H3 cookbook
- vLLM-Omni MiniMax H3 recipe
- Hugging Face diffusers MiniMax H3 documentation
- MiniMax official H3 integration catalog
- fal MiniMax H3 launch page and demos
- fal H3 text-to-video endpoint and pricing
- fal H3 image-to-video endpoint and pricing
- fal H3 reference-to-video limits and pricing
October: Single-GPU Streaming Is a Separate Experiment
ComfyUI's October 8 report describes ComfyStreamerH3 on RTX 5090 32GB at 448×256, combining community FastH3 weights and other optimizations. It does not establish real-time 768P or 2K on that card, nor the release of every official hosted component. See the setup and limitation guide. The current official API price table also separates output, reference inputs and 768P-to-2K regeneration; include all applicable charges.