MiniMax H3: Specs, H3 Max, Local Runs and Prompting Guide

MiniMax H3 is the open-weights video model from MiniMax, the Shanghai company behind the Hailuo AI app. It launched on July 31, 2026. It takes text, images, video and audio as input in one prompt and outputs video with native stereo sound, up to 15 seconds long and up to 2K resolution.

Quick answer: MiniMax H3 is one of the strongest video models you can use right now. On Artificial Analysis's image-to-video-with-audio leaderboard, it tops the board but is statistically tied with the rest of the leading group. You can use it in the Hailuo app, through MiniMax's API and partner platforms, or run the open weights yourself at 768p. MiniMax H3 is coming soon to CinemaDrop.

This guide, written for filmmakers, covers specs, H3 vs H3 Max, quality, access, local runs, prompting and workflow.

What Is MiniMax H3?

MiniMax calls H3 a “general-purpose multimodal generation model.” Earlier video models split the work into separate tools for text-to-video, image-to-video, first and last frames, reference shots and editing. H3 handles all of these in one model. You describe in plain language how your inputs relate to the shot you want. In MiniMax’s launch post, the example prompt asks the model to copy a Hitchcock camera move from one video, make the character from an image sing, and match the vocals to an audio clip.

“Hailuo” is MiniMax’s app brand, so some sites call the model Hailuo H3. They mean the same model. MiniMax’s own name for it is MiniMax H3.

Key facts from MiniMax’s official model card:

• Output: 4 to 15 seconds, 24 fps, with 32 kHz stereo audio.
• Resolution: the short side is 768 pixels by default. 2K comes from a separate H3-Regenerate-2K step.
• Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, among others.
• Dialogue: stable support for 11 languages, including English, Spanish, Japanese and Chinese.
• References: up to 9 images, 3 video clips and 3 audio clips, with no more than 12 files in total.
• Size: a 33B-parameter transformer that uses Qwen3-VL-32B as its text and vision encoder.

MiniMax pitches it for ads, product design, gaming and film titles.

MiniMax H3 vs H3 Max

“MiniMax H3 Max” draws about 720 searches a month, and few guides explain what it is. H3 Max is a separate listing of the model that has run on hosted platforms since August 2026. Higgsfield added it on August 31 with start and end frames at 768p. Artificial Analysis ranks it separately from the 768p base model. Here is how the two compare, using only published figures:

Prices are the Artificial Analysis list figures, read on Oct 24, 2026. What you actually pay depends on the platform. If you need 2K, use the base H3 through MiniMax’s own API. If your platform offers Max and you are working at 768p, it is a reasonable choice, but on the current leaderboard it ranks below the base model (see below).

How Good Is MiniMax H3?

The fairest independent read comes from blind human votes. On the Artificial Analysis image-to-video leaderboard (with audio, v1.1), MiniMax H3 (768p) tops the board at 1140 ±16 from 4,209 appearances (read Oct 24, 2026). The margins are tiny, though. The next model sits just two points behind at 1138 ±15, and the confidence intervals of the top six all overlap. H3 Max ranks sixth at 1125 ±17, 15 points behind the base model. H3 tops the board but is statistically tied with the leading group, so treat it as one of a handful of top models, not a clear winner.

The board was rebuilt on Oct 18, so older quoted scores don’t carry over.

Here is how the strengths and weaknesses break down for filmmakers:

• Follows references closely. This is the main reason the image-to-video score is so strong. Your start frame’s character, wardrobe and light carry through the shot.
• Multi-shot by design. MiniMax lists “native multi-shot modeling” in H3’s training, and its official prompts use timed cuts such as “[Shot 2] At 00:04.500, the camera cuts to…”.
• Dialogue and sound in one pass. Spoken lines, sound effects and music come out together, in stereo. In our dialogue test the lines and ambience were usable, but the lip movement looked stiff and generic, so plan retakes or a lip-sync pass for close-up dialogue.
• Limits: local runs top out at 768p. Creator tests describe a slightly glossy look. MiniMax itself says visual detail “can still be improved in certain scenarios.”

Medium two-shot, static camera, 16:9, photoreal 35mm film look. Night, a retro American diner booth: warm overhead pendant light, rain streaking the window behind them, neon glow outside. A tired waitress in her fifties sits across from a young man in a denim jacket, coffee cups between them. [00:00] She leans in and says, quietly: "You came all this way just to tell me that?" [00:05] He looks down at his cup, then back up, and says: "I came all this way because you never called." Soft diner ambience, rain on glass, a coffee cup set down on a saucer. Natural lip sync, realistic faces, no music, no on-screen text.

Where to Use MiniMax H3

You can reach H3 through several routes. They differ in resolution, price and how much setup they need:

A license note for studios

The open weights come under the MiniMax H3 Community License, not a standard open-source license. We checked the actual license text. It excludes four territories: the European Union, the United Kingdom, the Republic of Korea and the United States. If you are based in one of them, you need a separate license, and MiniMax invites people there to contact it about one. The license also requires written permission from MiniMax if your commercial products earn more than US$20 million a year. Commercial products must also display “MiniMax H3” in their interface. The license is governed by Hong Kong law. Hosted routes such as the Hailuo app and the API run under each provider’s own terms instead.

Can You Run MiniMax H3 Locally?

Yes. MiniMax has released two task checkpoints on Hugging Face. FL2VA covers text-to-video and first/last-frame video. Ref2VA covers reference-to-video with images, video and audio. Both output video with audio at 768p. Two parts of the full system aren’t in the open release:

• H3-Context-IR, the hosted step that turns your free-form prompt and references into the structured description H3-Base expects. MiniMax “strongly recommends” it, offers it as an API, and publishes a prompting guide so you can build your own.
• H3-Regenerate-2K, the 2K upscaling step. It is API-only for now.

For the MiniMax H3 ComfyUI setup, the model card links to official Comfy templates for reference-to-video and text-to-video. SGLang, vLLM and diffusers are also supported. If you are comparing open models in general, see our guide to the best open-source AI video generators.

Hardware: the Oxen.ai team, writing from hands-on runs, suggests at least 24 GB of VRAM. With a community NF4 quantization and offloading, H3 fits in as little as 8 GB. For higher-precision weights and longer clips, they put the sweet spot at 50–60 GB, split across GPUs.

Prompting MiniMax H3 for Film Shots

H3 responds best to the structured format MiniMax uses in its examples. That format has three parts: a shot-by-shot visual description with timestamps, an overall_soundscape line, and a non_diegetic_music line. For a film, that maps neatly onto a shot list. For more general prompt craft, see our AI video prompt guide.

[Shot 1] Live-action, [framing], [camera move]. [Setting, light, time of day]. [Character: age, look, wardrobe] [action].
[Shot 2] At 00:05.000, the camera cuts to [framing] of [character]. They say in a [tone] voice (S1): [English] "[line]."
overall_soundscape: [room tone, key effects, when they hit]
non_diegetic_music: [genre, instruments, tempo, where it swells or stops]

Three habits help most:

1. Give every reference one job. For example: “Image 1 is the character, Video 1 is the camera move, Audio 1 is the voice.” H3 reads those relationships from your words.
2. Time your cuts. Timestamp each new shot, and keep the sequence within 15 seconds.
3. Script the sound. Write the dialogue, room tone and music cue, or H3 invents them.

Slow, steady push-in toward the woman in the red coat. Light rain falls; reflections ripple on the wet pavement. After a beat she turns her head and shoulders to look straight into the camera, calm and searching. Keep her face, coat, the street and the blue-hour light exactly as in the image. Sound: soft rain, distant traffic hiss, her footsteps stop on the wet stone. No dialogue, no music, no text.

[Shot 1] 00:00, wide shot: an elderly lighthouse keeper in an oilskin coat climbs a narrow iron spiral staircase inside a stone lighthouse at dusk, lantern in hand, wind howling outside. [Shot 2] At 00:05, the camera cuts to a close-up of his weathered face and hands gripping the cold rail, breath visible, determined. [Shot 3] At 00:10, the camera cuts to the great lamp at the top: he throws the switch, the lamp ignites and the Fresnel lens begins to turn, light sweeping out over a stormy sea. Photoreal, 16:9, natural film grain. Sound: boots on iron steps, wind, a heavy switch clunk, the hum of the lamp and a distant foghorn. No music, no text.

Using MiniMax H3 in a Film Workflow

H3 makes 15-second shots, and a film needs dozens that match. Lock the look before you generate any motion:

1. Script. Break the scene into shots with dialogue.
2. Elements. Define each character, location and prop once, with reference images.
3. Storyboard. Generate a still for every shot from those Elements, so continuity is solved in images first.
4. Shots. Animate each storyboard frame with image-to-video. This is where H3’s close reference-following pays off.

CinemaDrop is built around exactly this flow. Its Script Wizard feeds @-taggable Elements, which feed an AI storyboard generator, and each board frame becomes a shot through storyboard image-to-video. MiniMax H3 is coming soon to CinemaDrop. Until then, you can run the same workflow today with the video models already in the app. See the image-to-video docs and the plans.

MiniMax H3 FAQ

How does MiniMax H3 work?

A hosted step called H3-Context-IR first turns your text and references into a detailed structured description. The 33B-parameter H3-Base model then generates video and stereo audio together at 768p. An optional H3-Regenerate-2K step re-runs the result with the original context to produce 2K. MiniMax says this recovers fine detail, such as small text, better than a standard upscaler.

Is MiniMax H3 any good?

Yes, it is one of the best available. It tops the Artificial Analysis image-to-video-with-audio board (1140 ±16) but is statistically tied with the leading group. Its standout strengths are following references and generating dialogue and sound in the same pass. In our test, lip movement on dialogue was stiff, so close-ups may need extra work.

What is the difference between MiniMax H3 and H3 Max?

H3 is the base model. It is open-weight at 768p and can produce 2K through MiniMax’s API. H3 Max is a hosted listing at 768p with start and end frames. On the Artificial Analysis board, the base H3 (768p) ranks first and H3 Max ranks sixth, though their confidence intervals overlap.

How can you use MiniMax H3 for free?

The open weights cost nothing to download, but they come with license terms and territory exclusions (see the license note above) and need a strong GPU. Hosted plans and free trials change often, so check the provider’s own page before you rely on one.

Can you run H3 locally, and what hardware do you need?

Yes, at 768p, using ComfyUI, diffusers, SGLang or vLLM. Plan on at least 24 GB of VRAM. Quantized community builds can fit in about 8 GB, and 50–60 GB is comfortable for higher precision and longer clips. We found no reliable source on AMD support.

Is MiniMax a Chinese company?

Yes. MiniMax’s Hong Kong listing documents name Shanghai as its head office. The company also runs the Hailuo AI app.

Is MiniMax H3 free or paid, and how much does it cost?

Hosted use is paid. Artificial Analysis lists MiniMax H3 (768p) at $4.80 per minute of video and H3 Max at $5.76 per minute. MiniMax says that at 2K, H3’s per-second price is “less than a third of mainstream models.” That is the vendor’s own claim, not an independent figure. Hailuo app plans change, so check the live pricing page.

Is MiniMax H3 available in CinemaDrop?

Not yet. H3 is coming soon to CinemaDrop. You can already plan the shots in CinemaDrop and animate them with the video models it offers today.