MiniMax H3 is the most capable open-weights-bound video generation model available through an API right now — it produces 15-second clips at native 2K resolution with synchronized stereo audio in a single pass, and it takes the top spot on the Artificial Analysis leaderboard for video editing. Released July 31, 2026, it undercuts competitors at roughly $7.80 per minute of 2K video compared to $20+ for Kling 3.0 and Seedance 2.0, and MiniMax has committed to publishing the model weights under a community license for organizations under $20M in revenue.
For content teams, marketers, and builders who have been stitching together a silent video model, a separate TTS engine, and a video editor just to ship a single clip, H3 collapses that entire pipeline into one API call. The model reads text, reference images, video clips, and audio inputs simultaneously, then outputs a finished shot — picture and sound together, already in sync.
Last verified: August 1, 2026
- Released July 31, 2026 via the MiniMax Open Platform
- Ranks #1 Video Editing, #2 Text-to-Video, #3 Image-to-Video on Artificial Analysis
- $0.13/second ($7.80/min) at 2K with audio — less than a third of mainstream competitors
- Open weights promised "in the coming days" under MiniMax Community License
- Pricing/limits/availability are volatile — verify before building a production pipeline
What Is MiniMax H3?
MiniMax H3 is a general-purpose omni-modal generation model built by MiniMax, a Shanghai-based AI lab. Unlike its predecessors in the Hailuo line (Hailuo 01, Hailuo 02, Hailuo 2.3), which were specialized video generators limited to silent 1080p output, H3 is a single unified model that handles text-to-video, image-to-video, reference-guided generation, and instruction-based editing — all with native stereo sound.
The model succeeds the Hailuo 2.3 series and is also known as Hailuo 3.0 on some platforms. The API model ID is MiniMax-H3.
How Does MiniMax H3 Compare to Hailuo 2.3?
The jump from Hailuo 2.3 to H3 is generational, not incremental. Here is what changed, verified against MiniMax's official announcement and platform documentation:
| Feature | Hailuo 2.3 | MiniMax H3 |
|---|---|---|
| Max resolution | 1080p | Native 2K (2560×1440) |
| Max clip length | ~10 seconds | 5–15 seconds (extendable to ~30s via Extend tool) |
| Audio | Silent | One-pass synchronized stereo (dialogue, effects, ambient) |
| Reference inputs | Single image (first-frame) | Up to 9 images, 3 video clips, 3 audio clips per request |
| Video editing | Not supported | Instruction-based natural-language editing of finished clips |
| Architecture | Dedicated video generator | Unified omni-modal transformer (text + image + video + audio) |
| Price | ~$0.02/sec at 1080p | $0.13/sec at 2K with audio / ~$0.05/sec at 768p |
The most important architectural change is that H3 fuses all modalities — text, image, video, and audio — at the pretraining stage rather than bolting audio on afterward. This is why the sound arrives already synced to the picture: the model generates both in the same forward pass, not as two separate outputs stitched together.
What Makes MiniMax H3 Different From Other AI Video Models?
Three capabilities set H3 apart from closed competitors like Google Veo 3.1, OpenAI Sora 2, and Kuaishou Kling 3.0:
1. Native stereo audio in every generation. Veo 3.1 also generates audio, but most competitors — including Sora 2, most Runway workflows, and the earlier Hailuo models — produce silent clips. With H3, dialogue, sound effects, ambient tone, and music are written alongside the picture in the same pass. This eliminates the entire post-production audio pipeline: you no longer need a separate TTS model, a sound library, or hours spent aligning audio to frames.
2. Instruction-based video editing. H3 can take a finished clip and modify it through natural language — swap a subject, change a background, add a character, or transfer motion from a reference clip — without regenerating from scratch. This is rare at any price point. As of August 2026, it is the capability that earned H3 the #1 ranking on Artificial Analysis's video editing leaderboard.
3. Open weights (planned). MiniMax has stated in its launch announcement that it plans to release H3's weights "in the coming days, subject to applicable laws and regulations" under the MiniMax Community License. This would permit free non-commercial use and commercial use for organizations under $20M annual revenue with attribution — making H3 the strongest open-weights video model available, ahead of the previous leader LTX-2.3. As of August 1, 2026, the weights have not been published; no Hugging Face repository or model card exists yet. Treat the open-weights claim as a promise, not a download.
How Much Does MiniMax H3 Cost?
MiniMax H3 is priced at $0.13 per second at 2K resolution with audio — approximately $7.80 per minute of generated video. At 768p, the price drops further: MiniMax states it is "less than half the price of mainstream models' 720p" at that resolution.
Here is how it stacks up against the closest closed competitors, sourced from Artificial Analysis and AlphaSignal's coverage:
| Model | Price per minute (2K with audio) | Max resolution | Max clip length | Native audio |
|---|---|---|---|---|
| MiniMax H3 | $7.80 | 2K (2560×1440) | 15 seconds | Yes (stereo) |
| Kling 3.0 (Kuaishou) | $20.16 | 1080p | ~10 seconds | Yes (Omni tier) |
| Seedance 2.0 (ByteDance) | $22.45 | 1080p | ~10 seconds | No |
| Veo 3.1 (Google) | $0.50/sec ($30/min) | 1080p | 8 seconds | Yes |
Pricing volatility flag: These are published API rates as of August 1, 2026. Video model pricing moves fast — always verify at the official MiniMax pricing page before committing to a budget.
Note: MiniMax's consumer video app, Hailuo AI, offers a free tier with limited trial credits and paid plans from $14.99 to $199.99/month. The free tier carries a watermark and gives the lowest queue priority. For production work, the API is the right surface.
What Can You Build With MiniMax H3?
Based on MiniMax's official documentation and the model's input specifications, here are the four core generation modes and practical use cases for each:
Text-to-Video (t2va)
Supply a text prompt and get a finished clip with audio. The prompt supports up to 7,000 characters. You specify duration (4–15 seconds in integer values), resolution (2K or 768p), and aspect ratio.
Practical use: Generate B-roll for explainer videos. A prompt like "A drone shot gliding over a fog-covered mountain ridge at sunrise, golden light breaking through clouds, smooth camera motion" produces a cinematic establishing shot with ambient environmental audio — no separate sound design needed.
Image-to-Video (i2va)
Supply a first frame image, an optional last frame image, and a text prompt. The model animates between the frames. Aspect ratio is inherited from the input image.
Practical use: Animate a product photo. Upload a static image of a product on a white background, prompt "the product slowly rotates on a glossy surface, studio lighting," and get a rotating product demo clip with the natural sound of a studio environment.
How to Use the MiniMax H3 API: Step-by-Step
The API is asynchronous and works in three steps: create a task, poll for completion, then download the video. Based on MiniMax's official API documentation:
Step 1: Get an API Key
Create an account at platform.minimax.io, navigate to the API keys section, and generate a key. Set it as an environment variable:
export MINIMAX_API_KEY="your-key-here"
Step 2: Submit a Text-to-Video Generation Request
Send a POST to https://api.minimax.io/v2/video_generation with your API key, specifying the model, prompt, duration, and resolution:
import requests, os
api_key = os.environ["MINIMAX_API_KEY"]
headers = {"Authorization": f"Bearer {api_key}"}
response = requests.post(
"https://api.minimax.io/v2/video_generation",
headers=headers,
json={
"model": "MiniMax-H3",
"content": [
{"type": "text", "text": "A barista pulling an espresso shot in a sunlit cafe, steam rising, warm golden light, close-up"},
],
"duration": 10,
"resolution": "2K",
"ratio": "16:9",
}
)
task_id = response.json()["task_id"]
The ratio parameter is required for text-to-video (use "16:9", "9:16", "1:1", "21:9", etc.). For image-to-video, ratio is always "adaptive" — the model inherits it from the input image.
Step 3: Poll for Completion
Query the task status periodically until it succeeds. The download URL is returned directly on success — no separate file retrieval step:
import time
while True:
result = requests.get(
f"https://api.minimax.io/v2/query/video_generation?task_id={task_id}",
headers=headers
).json()
if result["status"] == "Success":
video_url = result["content"]["url"]
break
elif result["status"] == "Failed":
raise Exception(f"Generation failed: {result.get('base_resp', {})}")
time.sleep(5)
Step 4: Download and Use
Fetch the video from the returned URL. The output includes both video and synchronized audio in the same file.
What Are the Limitations of MiniMax H3?
Being honest about the constraints matters more than hype. Here is what H3 cannot do as of August 2026:
Weights are not published yet. MiniMax has announced plans to open the weights, but as of August 1, 2026, there is no Hugging Face repository or model card for H3. Self-hosting is not possible yet. All access is through the API or the Hailuo AI consumer app. (Vendor claim — MiniMax announcement, July 31, 2026)
License has a revenue cap. The planned MiniMax Community License permits commercial use only for organizations under $20M annual revenue, with attribution required. It is a source-available license with a cap, not a true MIT/Apache open-source license. For most independent creators and small studios the cap is irrelevant, but for any company integrating H3 at scale, reading the final license text when it publishes is essential. (Reported — South China Morning Post)
Clip length tops out at 15 seconds. A single H3 generation runs 5–15 seconds. For longer content, you stitch multiple clips — which means matching cuts and re-establishing the subject at each seam. Competitors like Seedance 2.5 (ByteDance) generate 30-second clips in one pass and can extend further.
No subscription billing yet. H3 is billed per-generation through the API. The video subscription packages on the MiniMax platform currently support only the Hailuo 2.3 and earlier models — H3 is not included in subscription plans yet.
One-third-party leaderboard ranking so far. The #1 video editing, #2 text-to-video, #3 image-to-video rankings come from Artificial Analysis, a single independent evaluation source. There is no official benchmark table or technical report from MiniMax yet. Treat rankings as a strong prior — run your own 3–5 shot test before committing a pipeline.
How Does MiniMax H3 Fit the Chinese AI Model Landscape?
MiniMax is one of several Chinese AI labs shipping frontier video models in 2026. Reuters reported that the H3 release "steps up competition in a fast-growing market led by rivals ByteDance" (Seedance) and Kuaishou (Kling). The broader pattern is that Chinese labs are shipping open-weights-bound models faster than their Western counterparts — a trend visible in the LLM space with Kimi K3 and in the agent infrastructure space with MiniMax M3's sparse attention architecture.
The strategic difference: Western closed models (Veo, Sora) publish polished product experiences but keep weights proprietary. Chinese labs like MiniMax ship through API + consumer app but commit to releasing weights, which enables self-hosting and fine-tuning — a different value proposition for builders who want control over their inference stack.
What This Means for You
If you build marketing videos or product demos: Start testing the H3 API now. The native audio + 2K output in a single $1.95/generation call collapses your video pipeline from "generate silent clip + add TTS + edit in Premiere" to "one API call." Run 3–5 generations on your actual prompts to validate quality before committing a production budget.
If you run an AI agent system: H3 can serve as the B-roll generation layer in an automated video pipeline. A script-generation agent can call the H3 API directly for scene-specific clips with matching audio, then an editing agent can assemble the final video — no human-piloted video editor required in the loop for the initial cut.
If you want to self-host: Wait for the weights. Audit the MiniMax Community License carefully when it publishes. Do not commit to a self-hosted architecture until the repository, model card, and license file all exist — announcements are not downloads.
If you are comparing AI video tools broadly: H3 leads on price-performance and video editing, but it is not the only option. See our coverage of new tools in 5 New AI Tools Launched in July 2026 and our analysis of AI video editing workflows for broader context on where video generation fits in a working AI stack.
FAQ
Q: Is MiniMax H3 the same as Hailuo 3.0?
A: Yes. H3 is the official MiniMax model name. Because previous generations were branded "Hailuo," the model is also listed as Hailuo 3.0 or Hailuo-03 on some platforms. They are the same model. The API model ID is MiniMax-H3.
Q: Does MiniMax H3 generate audio natively? A: Yes. H3 generates synchronized stereo audio — dialogue, sound effects, ambient tone, and music — in the same pass as the video. Both audio and picture are produced by a single forward pass through the model, not as separate outputs.
Q: Is MiniMax H3 open source? A: Not yet. MiniMax announced plans to release the weights "in the coming days" under the MiniMax Community License, which permits free non-commercial use and commercial use for organizations under $20M revenue with attribution. As of August 1, 2026, no Hugging Face repository or model card exists. It is a planned release, not a current one.
Q: How much does a MiniMax H3 video cost to generate? A: $0.13 per second at 2K resolution with audio — approximately $7.80 per minute. A full 15-second 2K clip costs about $1.95. At 768p, the price drops further. This is less than a third of competitors like Kling 3.0 ($20.16/min) and Seedance 2.0 ($22.45/min) at 1080p, per Artificial Analysis pricing data.
Q: How long can a MiniMax H3 video be? A: Between 5 and 15 seconds per generation, specified in integer seconds only. This is a 50% increase over Hailuo 2.3, which topped out at 10 seconds at 1080p. For longer sequences, multiple clips must be generated and stitched together.
Q: Where can I access MiniMax H3? A: Three channels: (1) the MiniMax Open Platform API at $0.13/sec for 2K with audio, (2) the Hailuo AI consumer app with free trial credits and paid plans from $14.99/month, and (3) third-party-hosted endpoints on platforms like fal.ai and OpenRouter.
Discussion
0 comments