Chinese AI firm MiniMax launched H3 on July 31, 2026 — a general-purpose multimodal video model that generates 15-second clips at 2K resolution with native stereo sound, at roughly one-third the per-second price of mainstream competitors. The company pledged to release open weights “in the coming days,” directly challenging the closed-source video AI models from ByteDance and Google.
🔍 THE BOTTOM LINE: H3 is the world’s strongest open-weight video AI model the moment its weights drop, according to benchmark platform Artificial Analysis. It leads in video editing, undercuts rivals on price by 2-3x, and runs on consumer hardware. The closed-source video generation market just lost its pricing power.
What H3 Actually Does
H3 is not a text-to-video model with add-ons. MiniMax describes it as a general-purpose multimodal generation model that reads text, images, video, and audio as one unified context and returns video with native stereo sound. The key specs: 2K output, 4-15 second clips, and audio generated in the same pass as the video — not bolted on afterward.
Previous video stacks split into separate expert models for text-to-video, image-to-video, first-and-last-frame, subject reference, motion reference, and video editing. H3 folds all of these into one model where reference and editing relationships are expressed through natural language. A prompt like “reference the camera movement from Video 1, have the character in Image 2 sing, match the vocals to Audio 3” works in a single call.
What is open-weight AI? Open-weight models release the underlying parameters — the mathematical values that encode an AI’s intelligence — so anyone can download, run, and modify the model on their own hardware. Unlike closed-source APIs (where you pay per use and never see the internals), open weights let developers build custom versions, run them locally, and avoid vendor lock-in. The catch: you need enough hardware to run the model yourself.
The Price War
SCMP reported, citing Artificial Analysis data, that H3 costs US$7.80 per minute at 2K resolution with audio. ByteDance’s Seedance 2.0 at 1080p costs US$22.45 per minute. Google’s Gemini Omni Flash undercuts both at US$6 per minute — but remains closed-source.
At 768p resolution, H3 is priced at roughly half of peer offerings. At 2K, it is about a third. Jefferies analysts confirmed the pricing gap in a note on July 31.
The competitive dynamics are sharpened by timing. ByteDance released its own Seedance 2.5 on the same day, boasting 30-second videos in a single run. ByteDance’s annual recurring revenue from large language models reached US$4 billion in July 2026, SCMP reported — making it China’s largest AI company by that metric. MiniMax is the challenger betting on openness.
Why Open Weights Matter Here
MiniMax’s previous Hailuo video models were proprietary. The decision to open H3’s weights — under the MiniMax Community License, which permits free non-commercial use and commercial use for organisations under US$20 million annual revenue — marks a strategic shift.
MiniMax said the decision was made to “support the open-source community, accelerate compatibility with a broader range of AI hardware, and make it easier for users to build their own customised versions.” Hardware compatibility was cited as a “key consideration since the earliest stages of H3’s design.”
The ComfyUI community confirmed day-zero support, reporting that H3 can run locally on a consumer RTX 3060 graphics card — a GPU that costs under $300. This matters because it removes the API dependency entirely for users with modest hardware. The open-weights video generation space has been thin: China’s broader open-weights strategy has produced strong language models like DeepSeek V4 and Qwen 3, but video generation has lagged behind text.
Where H3 Falls Short
H3 is not the best at everything. Artificial Analysis ranked it as the world’s leading model for video editing specifically, but it trailed Google’s Gemini Omni Flash in text-to-video tasks and ranked behind both ByteDance’s Seedance 2.0 and Gemini Omni Flash in image-to-video capabilities. The competitive landscape is fragmented by task type — no single model wins everywhere.
The open-weights release is also not yet complete. As of August 3, the weights were “slated for release in the coming days” but not yet available for download. The API and consumer-facing Hailuo AI platform are live; local deployment is not.
The Bigger Picture
MiniMax’s pricing and open-weights move fits a pattern in Chinese AI: margin compression at the frontier where companies use aggressive pricing and openness to capture market share from established players. The edge AI hardware market is growing in parallel — as more devices run AI locally, the demand for open-weight models that can run on consumer hardware will increase.
For the creative industries — advertising, e-commerce, gaming, film pre-visualisation — H3’s pricing could make AI video generation a routine tool rather than a premium service. At $7.80 per minute of 2K output with native audio, the cost of producing a 15-second product video drops to under $2. That changes the economics for small studios and independent creators who previously could not justify the per-clip cost of closed-source alternatives.
❓ FAQ
What makes MiniMax H3 different from other video AI models? H3 handles text, images, video, and audio as a single unified input — you can ask it to reference a camera movement from one video, a character from an image, and vocals from an audio clip, all in one prompt. It also generates native stereo audio in the same pass as the video, rather than adding sound afterward.
When can I download the open weights? MiniMax said “in the coming days” as of July 31. The API and consumer platform are live now. The license allows free non-commercial use and commercial use for organisations under US$20 million annual revenue.
Can H3 run on consumer hardware? Yes. The ComfyUI community reported day-zero support and confirmed H3 can run locally on an RTX 3060. This is a significant shift — most high-end video AI models require data centre GPUs.
How does H3 compare to ByteDance’s Seedance? H3 leads in video editing and costs roughly a third of Seedance’s price at 2K. Seedance 2.5 generates longer clips (30 seconds vs 15) and ranks higher in image-to-video. No single model wins every category.
🔍 THE BOTTOM LINE: MiniMax H3 is the first open-weight video model that is both competitively priced and capable of running on consumer hardware. The video AI market has been a duopoly of closed-source products from ByteDance and Google. H3 doesn’t need to win every benchmark — it just needs to be good enough, cheap enough, and open enough to break the pricing power of the incumbents.
