Video · MiniMax

MiniMax H3 (Hailuo)

MiniMax's 2026 Hailuo models with stereo sound. H3 Max runs in private mode at a low price. Pick it for private, affordable clips.

Private, some versions anonymized Low cost
Latest versionMiniMax H3 Max RecastOctober 2026
Artificial Analysis Image to Video Arena1,195 Elo#1 of 22 on Artificial Analysis Image to Video Arena
Price$0.0279 per second of video$0.14 a 5 s clip
Key specClips up to 15 s, with sound

What sets it apart

MiniMax H3 is an omni-modal model: text, images, video and audio go in, and video with stereo sound comes out, up to 2K and 15 s with native multi-shot. Its weights are open, and its Max tier leads the Artificial Analysis image-to-video arena at the time of reading.

  • Open weights, 33B parameters
  • Stereo sound, up to 2K, 15 s
  • Max tier adds camera keyframes

Recommended for: keep a video project private.

Strengths and limits

Where it shines

  • H3 Max and Turbo run in private mode (no data kept)
  • Turbo is among the cheapest clips with sound
  • Multi-Angle makes product spins and bullet-time from one photo

Think twice when

  • H3 Max stops at 1080p; 2K needs the anonymized H3
  • Generated clips start at 5 s; Recast follows the length of your source video
  • 5 to 15 s, native sound (Recast keeps the source audio)
  • H3 Max: 480p to 1080p; H3: 768p or 2K
  • Max versions private, plain H3 anonymized

Use cases

Orbit around a product photo

Using the Multi-Angle mode on my sneaker photo: keep the sneaker frozen and orbit the camera 90 degrees to the left, then rise slightly. 6 seconds.

The Max camera mode moves the viewpoint while the scene stays still.

Character from reference images

Use my 3 reference images of the explorer and my 5 second voice clip. She reads a map in a jungle and says 'We are close.'

It takes up to 9 images, 3 videos and 3 audio clips.

Cheap 768p preview

A paper boat drifting down a rainy gutter, close-up, 8 seconds, gentle rain sound.

768p costs well under 2K, good for testing.

Against its rivals

  • Wan 3.0Wan 3.0 runs to 30 s but has closed weights.#3 of 22 on Artificial Analysis Image to Video Arena (1,164 Elo)
  • Seedance 2.5Seedance 2.5 reaches 30 s with a larger reference budget.#2 of 18 on Artificial Analysis Text to Video Arena with audio (1,142 Elo)
  • FLUX 3 videoFLUX 3 offers 20 s clips with keyframes and a draft mode.#5 of 18 on Artificial Analysis Text to Video Arena with audio (1,127 Elo)
Compare side by side ↗

Version history

Each version, dated, with what it changed over the one before.

  1. Hailuo 02Earlier, not in Sub Rosa
    • Native 1080p, triple the parameters and four times the training data of its predecessor.
    • 768p at 6 or 10 s, 1080p at 6 s, no audio.
    Source : minimax.io
  2. Hailuo 2.3Earlier, not in Sub Rosa
    • Better body motion, stylization (anime, ink wash) and facial micro-expressions.
    • Adds a cheaper Fast model at the same price tier.
    Source : minimax.io
  3. MiniMax H3
    • General-purpose omni-modal model: text, image, video and audio in, one shared framework.
    • Video with native stereo audio up to 2K and 15 s, with native multi-shot.
    • Open weights on Hugging Face under the MiniMax H3 Community License.
    Source : minimax.io
  4. MiniMax H3 Max RecastLatest
    • Video to video: replaces people in an existing clip with identities from reference photos, keeping the source motion, camera work, cuts and audio
    • Renders at 768P or 1080P
    Source : eachlabs.ai

Benchmarks

How to read these scores: what a benchmark tells you.

Artificial Analysis Image to Video Arena: #1 of 22 models measured in the catalog. The same blind duel, starting from one image: which clip brings it to life better. An Elo rating: higher is preferred more often. 100 points apart means the higher-rated model wins about 64% of duels.

Artificial Analysis Image to Video ArenaElo rating · Artificial Analysis, October 4, 2026
  1. 1. MiniMax H3 MaxMiniMax H3 (Hailuo)1,195
  2. 2. MiniMax H3MiniMax H3 (Hailuo)1,181
  3. 3. Wan 3.0Wan 3.01,164
  4. 4. HappyHorse 1.1HappyHorse1,104
  5. 5. Grok Imagine 1.5Grok Imagine video1,098
  6. 6. HappyHorse 1.0HappyHorse1,083
  7. 7. Veo 3.1 Full QualityVeo 3 and 3.11,082
  8. 8. Wan 2.7Wan 2.x1,077
  9. 9. Grok ImagineGrok Imagine video1,072
  10. 10. Veo 3.1 FastVeo 3 and 3.11,066
See the numbers
ModelElo
1. MiniMax H3 Max1,195
2. MiniMax H31,181
3. Wan 3.01,164
4. HappyHorse 1.11,104
5. Grok Imagine 1.51,098
6. HappyHorse 1.01,083
7. Veo 3.1 Full Quality1,082
8. Wan 2.71,077
9. Grok Imagine1,072
10. Veo 3.1 Fast1,066

Artificial Analysis Text to Video Arena with audio: #3 of 18 models measured in the catalog. The duel for clips with sound: picture, voices, effects and how well they match. An Elo rating: higher is preferred more often. Only models that make their own sound take part.

Artificial Analysis Text to Video Arena with audioElo rating · Artificial Analysis, October 4, 2026
  1. 1. Wan 3.0Wan 3.01,156
  2. 2. Seedance 2.5Seedance 2.51,142
  3. 3. MiniMax H3MiniMax H3 (Hailuo)1,138
  4. 4. MiniMax H3 MaxMiniMax H3 (Hailuo)1,134
  5. 5. Flux 3FLUX 3 video1,127
  6. 6. Seedance 2.0Seedance 2.0 and 1.51,119
  7. 7. Gemini Omni Flash 1.1Gemini Omni Flash1,113
  8. 8. Grok Imagine 1.5Grok Imagine video1,044
  9. 9. HappyHorse 1.1HappyHorse1,043
  10. 10. Wan 2.7Wan 2.x1,029
See the numbers
ModelElo
1. Wan 3.01,156
2. Seedance 2.51,142
3. MiniMax H31,138
4. MiniMax H3 Max1,134
5. Flux 31,127
6. Seedance 2.01,119
7. Gemini Omni Flash 1.11,113
8. Grok Imagine 1.51,044
9. HappyHorse 1.11,043
10. Wan 2.71,029

Artificial Analysis Text to Video Arena: #2 of 5 models measured in the catalog. People watch two silent clips made from the same description, without knowing the models, and pick the better one. An Elo rating: higher is preferred more often. 100 points apart means the higher-rated model wins about 64% of duels.

Artificial Analysis Text to Video ArenaElo rating · Artificial Analysis, October 4, 2026
  1. 1. Wan 3.0Wan 3.01,129
  2. 2. MiniMax H3MiniMax H3 (Hailuo)1,118
  3. 3. MiniMax H3 MaxMiniMax H3 (Hailuo)1,115
  4. 4. Seedance 2.5Seedance 2.51,102
  5. 5. Flux 3FLUX 3 video1,094
See the numbers
ModelElo
1. Wan 3.01,129
2. MiniMax H31,118
3. MiniMax H3 Max1,115
4. Seedance 2.51,102
5. Flux 31,094

All 12 models in this family

ModelUsePrivacyPrice
MiniMax H3minimax-h3-text-to-videoText to videoAnonymized$0.0728 per second of video$0.364 a 5 s clip
MiniMax H3minimax-h3-image-to-videoImage to videoAnonymized$0.0728 per second of video$0.364 a 5 s clip
MiniMax H3 R2Vminimax-h3-reference-to-videoReferences to videoAnonymized$0.0729 per second of video$0.365 a 5 s clip
MiniMax H3 Enhancedminimax-h3-enhanced-text-to-videoText to videoAnonymized$0.0728 per second of video$0.364 a 5 s clip
MiniMax H3 R2V Enhancedminimax-h3-enhanced-reference-to-videoReferences to videoAnonymized$0.0729 per second of video$0.365 a 5 s clip
MiniMax H3 Maxminimax-h3-max-text-to-videoText to videoPrivate$0.0279 per second of video$0.14 a 5 s clip
MiniMax H3 Maxminimax-h3-max-image-to-videoImage to videoPrivate$0.0279 per second of video$0.14 a 5 s clip
MiniMax H3 Max R2Vminimax-h3-max-reference-to-videoReferences to videoPrivate$0.0432 per second of video$0.216 a 5 s clip
MiniMax H3 Max Multi-Angleminimax-h3-max-multi-angleImage to videoPrivate$0.018 per second of video$0.09 a 5 s clip
MiniMax H3 Max Turbominimax-h3-max-turbo-text-to-videoText to videoPrivate$0.0144 per second of video$0.072 a 5 s clip
MiniMax H3 Max Turbominimax-h3-max-turbo-image-to-videoImage to videoPrivate$0.0144 per second of video$0.072 a 5 s clip
MiniMax H3 Max Recastminimax-h3-max-recastText to videoPrivateShown in the app

Specifications

SpecificationMiniMax H3
ReleasedJuly 31, 2026
Parameters33B
WeightsOpen, MiniMax H3 Community License
Acceptstext, images, video, audio
Producesvideo, audio
Clip length5 to 15 s, with sound
Resolutions768P, 2K
Price$0.0728 per second of video
PrivacyAnonymized

Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.