Aug 3, 2026, 10:09 AMyear 2026relevance 28023 min read

MiniMax H3 · 2K Video, Native Audio and Multimodal References

The HitPaw Edimakor channel presents MiniMax H3 as a model capable of understanding text, image, video and audio within the same creative context. The proposal is real, but the less than three-minute demo omits important differences between the 2K API, the open base model, the territorial license and the integration announced for the publisher. This guide confirms specs, calculates costs, and turns launch into a testing workflow for filmmakers, brands, and creative teams.

Video credit: HitPaw Edimakor.Source video: MiniMax H3: The Multimodal AI That Gets Your Whole Idea |One Model for Text, Images, Video and Audio.

Quick Summary

MiniMax launched H3 on July 31, 2026 as a multimodal generative system. The official API accepts text, images, videos and audio as context, generates 4-15 second clips in 768p or 2K and produces native stereo audio along with the video.
The HitPaw Edimakor video is a promotional presentation, not an independent test. Does not show complete prompts, seeds, costs, failed attempts, controlled comparison or actual usage within Edimakor. The applications presented should be read as possibilities, not as guaranteed quality.
Reference mode accepts up to nine images, three videos and three audio clips, with a maximum of 12 files. Reference video and audio have limits of 2 to 15 seconds, and audio cannot be used alone without at least one image or video.
The expression 2K needs context. The API offers 2K, but the published base model typically works with a short side of 768 pixels. The full 2K pipeline uses H3-Context-IR and H3-Regenerate-2K, hosted modules that are not yet a full part of the open version.
The weights were published under the MiniMax H3 Community License, not under a permissive license like Apache or MIT. The license excludes the EU, United Kingdom, South Korea and the United States from the authorized territory. In Portugal, it should not be assumed that executing or distributing weights is permitted without separate authorization.
The video announces MiniMax H3 on HitPaw Edimakor for the following week. As of August 3, 2026, the publisher's public pages still mention Hailuo generically and do not show H3 in the credits table. The integration should be treated as announced, not yet publicly confirmed as available.

--what to try first

Credit the HitPaw Edimakor channel for the embedded presentation and MiniMax for the template and technical documentation.
Use multimodal as a relationship between references, not just as a list of uploaded files.
Assign an explicit role to each input: identity, style, movement, camera, voice, rhythm or initial and final frame.
H3 generates between 4 and 15 seconds per task. Larger sequences continue to require planning, continuity, editing and final mixing.
Native stereo audio means dialogue, music and effects can be modeled alongside the image. It does not guarantee artifact-free final sound or a ready-to-deliver mix.
In the MiniMax API, 15 seconds costs $1.95 at 2K or $1.20 at 768p, before additional video and image references.
The first five reference images are free in the API. Each additional image costs $0.04, and video referrals are charged per second.
Don't call H3 open source without explaining the license. The more accurate formulation is open weights under a restrictive community license.
In the EU, including Portugal, do not perform, redistribute or offer weights based solely on the download button. Review the license and look for appropriate authorization.
The 2K from the official service is not equivalent to the local base output. High-resolution traversal again uses context and hosted regeneration to recover detail.
ComfyUI's native support is useful for studying T2V, I2V, and R2V, but the local workflow requires large downloads, hardware resources, and inference time.
The Edimakor integration was announced by the channel itself, but confirm the model selector, resolution, duration and credit consumption within the installed version.
Test first in 768p, validate composition, identity, movement and audio, and only then pay for the 2K generation.
Compare at least four seeds or repetitions per prompt and record success rate, cost per approved clip, and continuity failures.
Don't use a promotional reel as a benchmark. Evaluate H3 with references and requirements from your own project.

Credit and research status

The embedded video was published by the official HitPaw Edimakor channel on July 31, 2026. The narration introduces the MiniMax H3, summarizes applications and announces future integration into the editor. The material serves as an introduction to the product, not as an independent review.

This article cross-references the transcript provided with the official MiniMax announcement, the V2 API, the pricing table, the weight repository, the license, the ComfyUI documentation and the Edimakor public pages consulted on August 3, 2026.

Specifications, prices, availability and terms can change quickly. Always check current documentation before starting production or purchasing credits.

The Short Answer: What Makes MiniMax H3 Different

H3 attempts to replace multiple specialized paths with a single context representation. Instead of choosing one model for text, another for image-to-video, another for motion reference and another for audio, the user describes how text, images, videos and sound relate to the desired output.

In practice, you can ask an image to control the character, a video to control the camera movement and an audio to control the voice or rhythm. The prompt stops describing just the visual result. It becomes a map of responsibilities between references.

This does not mean that H3 is a universal chatbot that delivers any final format with the same interface. The public product analyzed is mainly a video generation and editing system with native audio, accessible via API, suppliers and local workflows.

Confirmed specifications: duration, resolution, audio and formats

The V2 API accepts clips from 4 to 15 seconds. Available resolutions are 768P and 2K. Text-to-video supports 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. In reference-to-video mode there is also adaptive, while image-to-video inherits the aspect ratio of the provided frames.

The model card indicates video at 24 fps and stereo audio at 32 kHz. Voice, effects and music are previewed together with the video and decoded to a single MP4, rather than being added by an external tool at the end.

Each order requires text. Even when there are images, video or audio, the prompt remains mandatory because it describes the relationship between the context and the target.

T2V, I2V, first-last frame and reference-to-video

Text-to-video uses text only. Image-to-video can receive a first frame, a last frame or both. Reference-to-video accepts a combination of materials that guide identity, style, movement, camera, voice and sound.

First-last frame and reference-to-video are different paths in the API. You cannot mix frames marked as first_frame or last_frame with inputs marked as reference_image, reference_video or reference_audio in the same request.

The published base model also separates fl2va weights, used by text-to-video and image-to-video, from ref2va weights, used in references mode. An interface may appear unified, but local workflows continue to require the correct variant.

How multimodal references work

The API allows for up to nine reference images, three videos, and three audio files, with a combined maximum of 12 files. Each image can be up to 30 MB in size. Each video can be up to 50 MB, and each audio up to 15 MB.

Reference video must be between 2 and 15 seconds, with a total maximum of 15 seconds. The same range and total apply to audio. A reference audio needs to accompany at least one image or video.

More references do not automatically mean better results. Start with a reference for responsibility. If two images dispute identity or two videos describe incompatible movements, the model has to resolve a conflict that the prompt should avoid.

Native stereo audio: what changes in the workflow

In an audiovisual model, mouth timing, impact, ambience and music can be planned in the same time window as the movement. This makes it possible to request synchronized dialogue, a physical sound in the right frame, and a musical change that accompanies a cut.

The advantage is temporal coherence. The risk is loss of separate control. An incorrect voice, strange noise or inappropriate music can also force you to regenerate an image that was good. For professional work, export and evaluate image, dialogue, effects and music as distinct conceptual layers, even when they arrive in the same file.

Reserve the final mix to an editor or DAW when you need loudness, cleanup, subtitles, music rights, or localized versions. Native audio reduces editing, but does not eliminate post-production.

Because 2K is not just an upscale

MiniMax describes H3-Regenerate-2K as a second pass. The 768p result and original context are fed back to the H3, which regenerates the high-resolution version and attempts to recover small text and detail that conventional super-resolution would have to guess.

The open base model defaults to a short side of 768 pixels. ComfyUI recommends around 1344 by 768 for the native canvas. H3-Context-IR and H3-Regenerate-2K are not fully published in the open version, so reproducing the official 2K output locally still requires services from the MiniMax platform.

This distinction is important for infrastructure planning. Having the H3-Base weights does not mean having the entire official 2K system offline.

How much does MiniMax H3 cost on API

As of August 3, 2026, the official chart charges $0.13 per second on 2K and $0.08 per second on 768p. A 15-second clip costs $1.95 at 2K or $1.20 at 768p, before the price of reference materials.

Reference audio is free. The first five images are free and each additional image costs $0.04. Reference video is charged per second at the same rate as the output resolution. A 15 second 2K order with 15 seconds of reference video can be as low as $3.90 before additional footage.

Regenerating 768p to 2K costs $0.05 per second of output, but the original materials are charged again. Set a per-shot budget and lock in identity, movement and timing at 768p before paying the final pass.

Open weights does not mean open license for everyone

The H3 weights were published on August 2nd under the MiniMax H3 Community License Agreement. It is a custom license. It is not MIT, Apache 2.0, or automatic approval for any country or product.

The license defines the applicable territory as the world minus the European Union, the United Kingdom, South Korea and the United States. It also restricts the use of outputs to improve other models, requires identification of AI-generated content in public environments, and imposes additional conditions on commercial products and services.

For a user in Portugal, the practical conclusion is simple: do not execute, distribute or offer weights based solely on the open tag. Seek authorization from MiniMax and appropriate advice. This note refers to the weights license and does not replace review of the separate terms of the hosted API. It is not legal advice.

Run H3 in ComfyUI

ComfyUI 0.30.0 or higher includes native templates for text-to-video, image-to-video, and reference-to-video. There are BF16, INT8 and a pruned INT8 variants, plus a Qwen3-VL-32B encoder and separate VAEs for video and audio.

Templates allow you to experiment with parameters, aspect ratio, duration and references without relying on a closed interface. However, the files are large and local inference can be slow. The documentation does not turn the configuration into a lightweight download for any laptop.

Before hardware, confirm the territorial license. Then choose a small preview, validate the workflow and only then increase resolution, duration or references.

MiniMax H3 on HitPaw Edimakor: announced, yet to be confirmed

The video claims that MiniMax H3 will arrive at HitPaw Edimakor the following week. As the publication took place on July 31, the announcement points to the beginning of August 2026.

As of August 3, the Edimakor generator's public page lists Hailuo among several suppliers, but does not identify H3. The credits page also shows Hailuo costs for 512p, 768p and 1080p, without a MiniMax H3 2K lineup. The July news page does not record the integration.

This does not prove delay. It just means that public availability was not yet documented on the pages consulted. When the option appears, confirm version, reference limit, audio, duration, resolution and credits. An integration may only expose part of the original API.

Film, ads, e-commerce and gaming: where H3 can help

For filmmaking, H3 is interesting in previs, title sequences, camera tests, performance with voice reference and variations of an existing shot. The 15 second limit fits better for shots and beats than for complete scenes.

In advertising and e-commerce, references can fix product, brand, movement and sound cue within the same order. Readable text and brand rendering are capabilities claimed by MiniMax, but must be tested with actual packaging, alphabets and sizes.

In gaming and UI, the model can prototype motion studies, loops, interfaces and conceptual kinematics. Generated outputs do not replace interactive assets, rigs, states or engine implementation.

How to Write a Multimodal Prompt for H3

Start with a sentence that describes the complete target. Then identify the role of each reference. Finish with a short timeline that separates shot, action, camera and audio.

A useful structure is: objective, reference map, mandatory continuity, temporal beats, camera, dialogue, effects, music and negative criteria. Don't let the model infer whether a video is for style or movement.

Keep a single dramatic conflict or main action per clip. A 15-second prompt can contain several shots, but each new cut increases the risk of identity, screen direction and audio losing continuity.

Testing protocol before using H3 in production

Choose three representative shots: a person with dialogue, a product with text, and a motion based on reference video. For each shot, prepare a minimal prompt and a multimodal version.

Generate at least four variants at 768p. Evaluate identity, pose, physics, camera, text, lip sync, audio, duration and background stability. Record cost and time until you get an approvable clip.

Only make 2K of successful candidates. Compare the regenerated version with 768p to confirm that details have improved without changing the face, letters, objects or timing. This comparison measures the real value of the 2K pass.

Verdict: a workflow change, not yet an entire production

MiniMax H3 makes the prompt more like direction: different materials are assigned roles and the model tries to resolve image, movement and sound together. This is the most relevant advance, not just the 2K number.

Launch also comes with complexity. Official 2K uses hosted modules, the cost grows with seconds and reference video, local execution does not include the entire pipeline, and the weights license excludes several important regions.

HitPaw Edimakor's video is a good introduction, but it anticipates integration and summarizes capabilities without testing failures. Use it to discover H3, then validate documentation, editor availability, and a small set of your own shots before changing the pipeline.

Practical prompts for MiniMax H3

These original prompts use the central principle of H3: each reference is assigned an explicit function. Adapt names and times to the mode available on your interface and only use materials that you can legally process.

Prompt pack credit: HitPaw Edimakor.

Text-to-video with synchronized audio

Tests a single shot, simple movement, and verifiable soundscape.

Target: continuous cinematic shot of {{duration between 4 and 15 seconds}}.

SCENE: {{place, time, subject and main action}}.
CAMERA: {{shot size}}, {{height and angle}}, movement {{physical trajectory}} at speed {{velocity}}. No cuts.
CONTINUITY: {{clothes, object, direction of gaze and fixed elements}}.
DIEGETIC AUDIO: {{ambiance}}, {{effect associated with action}}.
DIALOGUE: {{character}} says in {{language}}: "{{short line}}". Natural lip movement and end of speech before the last second.
MUSIC: {{style and intensity}}, low under dialogue.
FINAL: end in {{final composition}} with stable subject and background.

Avoid: unsolicited cuts, digital zoom, morphing, random text, overlapping voices and music covering dialogue.

Reference-to-video with explicit roles

It assigns identity, movement, camera and voice without creating conflicts.

Creates a video of {{length}} seconds in {{aspect ratio}}.

REFERENCE MAP:
- <Picture 1>: only controls the facial identity and hair of {{character}}.
- <Picture 2>: controls clothing, materials and palette.
- <Video 1>: only controls body movement and timing. Do not copy identity, clothing or scenery.
- <Video 2>: only controls camera trajectory and speed.
- <Audio 1>: controls only voice timbre.
- <Audio 2>: controls musical rhythm, without copying lyrics or protected melody.

TARGET: {{full shot description}}.
ACTION: {{single and verifiable action}}.
CAMERA: {{movement and final framing}}.
DIALOGUE: {{short line}}.
SOUND: {{ambiance and effects}}.

Preserve identity, clothing and screen direction. Resolve any ambiguity in favor of the functions defined in the map.

15 second multi-shot sequence

Divide an idea into three beats with planned continuity and audio.

Create a 15-second sequence with three shots and a single character.

GLOBAL CONTINUITY: {{identity, clothing, object, location, light and screen direction}}.

0.00-5.00 s, SHOT 1:
{{establishing shot, action, camera and sound}}.

5.00-10.00 s, SHOT 2:
{{medium or close-up}}, cut in continuous motion. {{action and dialogue}}.

10.00-15.00 s, SHOT 3:
{{reveal or payoff}}, maintain relative position and end in {{final frame}}.

AUDIO: continuous environment between cuts; synchronized effects; music with a single rise in reveal; non-overlapping dialogue.

Do not introduce new characters, props, brands or clothing changes.

Product reveal with branded text

Evaluates product preservation, readability and sound design.

Creates a product reveal of {{duration}} seconds.

<Picture 1> defines the shape, color, materials and logo of the product. Do not redesign.
<Picture 2> defines the visual environment and lighting, not the product.
<Video 1> only sets the turntable movement and camera speed.

SHOT: {{framing}} on {{surface and background}}.
ACTION: the product {{simple action}} while the camera {{movement}}.
TEXT IN FRAME: write exactly "{{short text}}" only once, with {{generic typography}}, without extra letters.
LIGHT: {{key, kidney and reflections}}.
AUDIO: {{stuff sound}}, {{short whoosh}}, {{music sting}}.
FINAL: product hero frame cleared during the last second.

Priority: product geometry, accurate text, stable logo and coherent reflections.

First-last frame with controlled transition

Plans the transformation between two frames without mixing the reference mode.

Uses the first given frame as the start and the last given frame as the destination.

DURATION: {{4 to 15 seconds}}.
TRANSFORMATION: {{what changes}}.
PRESERVE: {{identity, object, axis, horizon and elements that do not change}}.
CAMERA: {{static or simple movement}}.

TIMELINE:
- start: keep the first frame readable for {{time}};
- means: execute {{physical and causal transition}} without arbitrary dissolution;
- end: reach the composition of the last frame until {{tempo}} and stabilize.

AUDIO: {{ambiance}}; {{effect}} grows during the transformation and ends exactly when the last frame stabilizes.

Avoid facial morphing, duplicate objects, cuts and changes to the final frame.

Evaluation matrix for four generations

Turns prints into a useful comparison for production.

Evaluates four MiniMax H3 outputs from the same prompt.

Shot objective: {{objective}}
Prompt: {{prompt}}
References and functions: {{map}}
Resolution: {{768p or 2K}}
Cost per attempt: {{value}}

For each output, assign 0 to 5 and explain in one sentence:
1. identity and consistency;
2. execution of the action;
3. camera trajectory;
4. physics and contact;
5. readability of text or product;
6. lip sync;
7. effects and environment;
8. music and mixing;
9. background stability;
10. usefulness without correction.

Selects the best candidate, identifies repairable defects in post-production and defects that require regeneration. Calculates total cost and cost per approvable output. Do not change the criteria after seeing the results.

Video Timestamps

Introduction00:00 What is MiniMax H300:19 Multimodal generation capabilities00:54 Reference-to-video01:07 Film, advertising, e-commerce and gaming01:20 MiniMax H3 announced for Edimakor02:14 Conclusion02:38

FAQ

What is MiniMax H3?

It is a multimodal generative system from MiniMax that understands context consisting of text, image, video and audio and generates video with native stereo audio.

How long can an H3 video be?

The official API accepts integer durations between 4 and 15 seconds. Larger clips need to be divided into shots and assembled later.

MiniMax H3 really generates in 2K?

Yes, the API offers 2K output. The official system uses regeneration with the original context. The open base model normally works at 768p and does not include all the 2K pipeline modules on its own.

Is the audio created at the same time as the video?

Yes. H3 predicts video and audio latents together and produces stereo audio, including voice, effects and music. This does not dispense with review and final mixing.

What references can I use?

Reference-to-video supports up to nine images, three videos and three audio clips, with a combined maximum of 12 files. Audio needs to accompany an image or video.

Can I use first frame, last frame and references in the same request?

No. The API treats image-to-video with first or last frame and reference-to-video as mutually exclusive modes.

How much does a MiniMax H3 video cost?

As of August 3, 2026, the API charges $0.13 per second on 2K and $0.08 per second on 768p. A 15-second clip costs $1.95 or $1.20 before referral fees.

Are there a cost for reference images and videos?

The first five images are free and each additional image costs $0.04. Audio is free. Video is charged per second of input at the rate of the chosen resolution.

Is MiniMax H3 open source?

Weights are available but under the MiniMax H3 Community License. The designation open weights is more accurate than open source because there are territorial and usage restrictions.

Can I run the MiniMax H3 weights in Portugal?

The published license excludes the European Union from the authorized territory. Do not assume that the download permits execution or distribution in Portugal. Contact MiniMax and seek appropriate advice.

Does MiniMax H3 work on ComfyUI?

Yes. ComfyUI 0.30.0 or higher provides templates for text-to-video, image-to-video and reference-to-video, with native stereo audio.

Does the full 2K pipeline work offline?

Not yet fully published. H3-Context-IR and H3-Regenerate-2K remain linked to MiniMax hosted services in the full official workflow.

Is MiniMax H3 now available on HitPaw Edimakor?

The channel announced the integration for the following week. On August 3, 2026, the public pages consulted still did not explicitly list H3. Confirm within the latest version of the editor.

H3 replaces a video editor?

No. It generates shots with image and sound, but editing, continuity, correction, mixing, subtitles, versions and deliverables continue to require an editing workflow.

What is the best way to test H3?

Start with a short shot at 768p, use one reference per function, generate four variants and evaluate with fixed criteria. Only submit successful candidates for 2K.

Is the HitPaw Edimakor video an independent review?

No. It is a promotional presentation of the product channel. Summarizes specifications and applications, but does not present controlled comparisons, complete costs, or failed attempts.