MiniMax H3: Open-Weight Multimodal Video at One-Third the Cost

Introduction

On July 31, 2026, MiniMax — a Shanghai-based AI company valued around $4 billion — officially released H3, the third generation in its Hailuo video model line. Three things make this launch significant: it is the first multimodal video model announced with an open-weight plan, its API pricing undercuts competitors by roughly 3x when normalized for output quality, and it shipped with Day-0 ComfyUI support.

Unlike previous Hailuo releases — which accepted a text prompt or a single image — H3 takes text, images, video, and audio simultaneously as input and returns 2K video with native stereo audio in a single generation pass. No separate upscale stage, no post-hoc audio dubbing. One API call, one result.

What H3 Is — Not Just "a Video Model"

The biggest difference between H3 and existing video models is how it handles multimodal input. H3 does not split text-to-video, image-to-video, and audio-to-video into separate pipelines. Instead, all inputs — up to 9 reference images, 3 video clips, and 3 audio clips — go into a single transformer and are processed as a unified context.

This unlocks complex creative tasks that were previously multi-step workflows. An example from the official blog: "Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3." One sentence describes the relationship between inputs, and H3 handles all cross-modal logic internally.

For developers, the implication is clear: the video generation API is no longer a pipeline of discrete steps. Instead of calling T2V → upscale → audio dubbing separately, everything happens in a single endpoint. Integration complexity drops, development time shrinks.

Pricing: Getting the Comparison Right

H3's published API price is $0.13 per second of native 2K video. Key competitors: Kling (Kuaishou) around $0.07/second at 1080p, Sora (OpenAI) around $0.10/second at 1080p, and Veo 3 (Google) up to $0.40/second at 4K. At first glance, Sora appears cheaper than H3 on raw price ($0.10 vs. $0.13). But that comparison is misleading — Sora outputs 1080p with no audio. H3 outputs native 2K with stereo audio in the same pass.

The "3x cheaper" framing comes from MiniMax itself: they claim H3 at 2K costs less than one-third of mainstream models when normalized for equivalent quality. While they do not name specific competitors, market data supports the direction of the claim. A 15-second product ad at 2K with H3 costs $1.95. Achieving comparable quality with Sora (1080p + upscale + separate audio generation) easily pushes past $4–5.

In other words: H3 is not the cheapest option in raw dollars. But on a cost-per-quality basis — resolution plus audio — the gap is real.

Open Weights — But No Weights Yet

H3 is marketed as "open-weight," but as of August 1, 2026 — one day after launch — the weights have not shipped. MiniMax promises to release them "in the coming days, subject to applicable laws and regulations." The planned license is the MiniMax Community License: commercial use permitted for organizations under $20 million in revenue, with attribution required.

This follows a familiar pattern in Chinese AI: announce "open" first, ship weights later. DeepSeek, Qwen, and Kimi have all done this. It is not necessarily a red flag — MiniMax has a history of releasing M-series models openly — but developers should be cautious about building production pipelines that depend on weights that do not yet exist.

On the positive side: MiniMax states that hardware compatibility was a design consideration from the earliest stages of H3's development, including support for Chinese-manufactured chips. This is not a model trained on a specific GPU cluster and then ported afterward — the self-hosting intent was baked into the architecture from day one.

ComfyUI Day-0 Support — Runs on an RTX 3060

On launch day, ComfyUI shipped version 0.30.0 with native H3 support. This is a significant milestone: new video models typically take weeks to months to get stable integration in open-source tools.

The ComfyUI team implemented three key optimizations to bring H3 to consumer hardware:

• Modulation weight pruning: Roughly 40% of H3's parameters are modulation weights — weights that adjust model behavior based on input. The ComfyUI team found they could replace these with a functionally equivalent lookup table, dramatically shrinking memory footprint with no output quality loss.
• INT8 convrot quantization: Custom quantization for convolution layers that preserves higher accuracy than standard linear INT8.
• Dynamic VRAM offloading: Automatically moves model segments between GPU RAM and system RAM during inference.

The result: a 66% memory reduction — from 123.6 GB (full precision) to 42.5 GB. Combined with offloading, H3 can run on an RTX 3060 (12 GB VRAM). A 2K multimodal video model running on a mid-range gaming card — unthinkable six months ago.

Technical Architecture: Four Pillars

While MiniMax has not yet published a full technical report, the launch blog describes four core technologies. The common thread: all of them reduce inference cost without sacrificing quality.

Contextual Omni Representation — a multimodal captioning pipeline that converts complex inputs (images, video, audio) into compressed language representations. Each source sample requires roughly 100K tokens of inference, distilled to an average of approximately 4K tokens. Language serves as the bridge between input context and target video, enabling cross-modal relationships to be described in natural language rather than technical parameters. For developers, this means prompts like "make character A move like video B, voice from audio C" actually work — no need to break it into separate steps.

H3-VAE — a rebuilt tokenizer achieving 4× compression over the Hailuo 02 architecture. The direct consequence: lower training and inference costs, plus native 2K output without a separate super-resolution stage. This is one reason H3 undercuts competitors on price despite higher output quality.

H3-Omni Transformer — a training architecture that separates understanding compute from generation compute. Multimodal context causes sequence-length variance to triple; this separation improves training throughput by nearly 30%.

In-Context Regeneration — instead of using a dedicated super-resolution module to upscale low-res output to 2K, H3 regenerates its own low-resolution output in-context. This approach recovers fine details — small text, brand logos — that traditional SR "can only guess at." This is why H3 is strong at text rendering and brand marks, a persistent weakness in most current video models.

Takeaways for Developers

• Simpler API integration: A single endpoint handles text + image + video + audio → video + stereo audio. No more 4–5 step pipelines.
• Competitive cost-per-quality: $0.13/second for native 2K with audio. Sora is cheaper in raw dollars but at lower quality and without sound — on a fair comparison, H3 leads on cost-per-quality.
• Open weights are coming — but not here yet: The API is the only option for immediate integration. Do not plan self-hosted production infrastructure around unreleased weights.
• ComfyUI ready today: Test locally on an RTX 3060. Workflows are available in the ComfyUI template library — download, load the model, run.
• It is a specialized model: H3 is a video generation model — it does not replace LLMs or act as a chatbot. It solves a narrow problem, but solves it well.

Conclusion

H3 marks a meaningful shift in video generation: true multimodality, dramatically lower cost-per-quality, and — if the open-weight promise materializes — the first truly open frontier video model. Day-0 ComfyUI support and optimizations enabling RTX 3060 inference show the open-source ecosystem is ready.

One question remains: when will the weights actually ship? In recent Chinese AI history, the gap between "open announcement" and "code on Hugging Face" ranges from days (DeepSeek) to weeks (Qwen). For MiniMax, the answer is still pending.

Content assisted by AI (Amy 🌸). Reviewed by the author.