MiniMax launches H3 for video apps and APIs

• MiniMax shipped H3 to Hailuo and the MiniMax API; day-zero partner posts put it on fal, Vercel AI Gateway, OpenRouter, Pika MCP, Runway, Venice, Krea, and Arena through MiniMax's fal note, Vercel's code example, MiniMax's OpenRouter note, MiniMax's Runway note, and Arena's Battle Mode post.
• The builder hook is its reference budget: fal says in its launch thread that H3 combines up to 9 images, 3 video clips, and 3 audio clips, while OpenRouter's parameter note exposes first/last-frame conditioning plus image, video, and audio reference URLs.
• Artificial Analysis ranked H3 No. 1 for video editing, No. 2 for text-to-video, and No. 3 for image-to-video; its evaluation thread lists 2K pricing at $0.13/sec.
• Open weights are a later step: MiniMax repeated “coming soon” in its Vercel partner note, and MiniMax's launch blog says weights are planned for the coming days, subject to applicable laws and regulations.

MiniMax's launch blog is the keeper: H3's full-modality captioning pipeline reportedly spends about 100K tokens of inference on most source material and distills it to roughly 4K tokens. The same post says a tokenizer overhaul gave H3 a 4x effective sequence-length gain, and that 2K output uses in-context regeneration instead of a dedicated super-resolution module. fal's reference-to-video endpoint turns that into an interface: 9 images, 3 videos, and 3 audio clips as ordered references.

H3 shipped like infrastructure: one model launch, many surfaces.

• Hailuo and the MiniMax API went live, according to MiniMax's launch post.
• Vercel AI Gateway showed H3 behind generateVideo() with model: 'minimax/minimax-h3' in its code example.
• OpenRouter listed H3 as live through its launch post, and OpenRouter's model page shows the slug minimax/hailuo-3.
• fal published text-to-video, image-to-video, and reference-to-video routes through its endpoint thread.
• Pika put H3 on Pika MCP in its demo post.
• MiniMax said H3 had landed with Krea, OpenRouter, fal, Runway, and Vercel through its Krea note, its OpenRouter note, its fal note, its Runway note, and its Vercel note.
• Venice made H3 available anonymously, according to Venice's post.
• Arena added H3 to Text-to-Video and Image-to-Video Battle Mode in Arena's post.

The reference budget is the part that makes H3 feel different from a normal text-to-video endpoint. The supported inputs from fal's launch thread and OpenRouter's parameter note are:

• Up to 9 reference images.
• Up to 3 reference video clips.
• Up to 3 reference audio clips.
• Up to 12 reference files total on OpenRouter.
• First-frame and last-frame conditioning through OpenRouter's Video API.
• Native stereo audio on every generation.

MiniMax's official post describes H3 as an omni-modal generation model that jointly understands text, images, video, and audio, then outputs up to 15 seconds of 2K video with native stereo sound.

TomLikesRobots' early test matched the same shape: 12 references, audio and video references up to 15 seconds, and 2K output in his first H3 generation.

The launch created a small model-slug tax.

• Vercel's example uses minimax/minimax-h3 inside generateVideo() in its AI Gateway post.
• OpenRouter's model page uses minimax/hailuo-3, lists pricing from $0.13/sec, and says the model is hosted by one provider.
• fal's text-to-video endpoint uses minimax/hailuo-03/text-to-video and renders 2K clips from 5 to 15 seconds.
• fal's reference-to-video endpoint exposes the 9-image, 3-video, 3-audio reference path.

OpenRouter's video generation docs describe the generic flow as an asynchronous API: submit a video job, poll for status, download content, or attach a callback_url for webhooks.

Artificial Analysis had the densest third-party snapshot in the launch evidence.

• Leaderboards: H3 ranked No. 1 in Video Editing, No. 2 in Text-to-Video, and No. 3 in Image-to-Video, according to the evaluation thread.
• Evaluation tier: Artificial Analysis said its test used the 2K tier in the same thread.
• Output shape: the thread described H3 as generating 5 to 15 second clips at 24fps with native audio.
• Listed price: 2K was $0.13/sec, or $7.80/min, while a 768p tier was listed as coming soon at $0.09/sec, or $5.40/min.
• Price comparison: Artificial Analysis put H3 below Dreamina Seedance 2.0 1080p at $22.45/min, HappyHorse-1.1 at $9.90/min, and Kling 3.0 1080p at $20.16/min, but above Google's Gemini Omni Flash at $6.00/min.

The same thread said MiniMax plans to release H3 under the MiniMax Community License, with commercial use allowed for organizations under $20M in revenue and prominent attribution requirements.

MiniMax's official launch blog names four implementation choices worth filing away.

• Contextual Omni Representation: MiniMax says most source material runs through about 100K tokens of full-modality inference, then gets distilled to roughly 4K tokens.
• H3-VAE: MiniMax says it rebuilt the tokenizer, gained 4x effective sequence length, and lowered training and inference cost enough to support native 2K output.
• H3-Omni Transformer: MiniMax says it set aside the Hailuo-02 architecture, separated understanding and generation workloads, and improved training throughput by nearly 30%.
• In-context Regeneration: MiniMax says H3 regenerates its own low-resolution output into 2K using the base model, instead of relying on a dedicated super-resolution module.

MiniMax also says H3 still needs better multimodal understanding, more scale, and improved visual detail in some scenes.

Open weights stayed future-tense across the launch.

• Vercel: “Open weights coming soon” in MiniMax's Vercel note.
• Krea: “Open weights coming soon” in MiniMax's Krea note.
• OpenRouter: “Open weights are coming soon” in MiniMax's OpenRouter note.
• fal: MiniMax tagged the launch #OpenWeights in its fal note.
• Runway: “Open weights coming soon” in MiniMax's Runway note.
• Arena: “Open weights dropping soon” in MiniMax's Arena note.

MiniMax's official blog gives the reason list: support the open-source community, accelerate compatibility with China-developed AI chips, and let users customize their own versions. The same post says hardware compatibility was considered from the earliest stages of H3's design.