MiniMax H3: How One Transformer Replaced a Whole Pipeline of Video Generation Models

Video generation has long been a patchwork of specialist models. You needed one model for text-to-video, another for image-to-video, a separate system for audio, and yet another for editing. MiniMax H3, released on July 31, 2026, takes a different approach: a single 33-billion-parameter transformer that reads text, images, video clips, and audio together, then outputs a 4–15 second video with native stereo audio — all in one pass.

The model is also the first commercially viable open-weight flagship video generation system, ranking #1 in video editing, #2 in text-to-video, and #3 in image-to-video on the Artificial Analysis leaderboard as of early August 2026.

The Problem H3 Is Solving

Before H3, the video generation landscape was fragmented by design. Image generation was split into separate expert models for text-to-image, editing, subject reference, motion reference, and style reference. Audio generation treated voice, sound effects, and music as independent domains. Video generation was further splintered into text-to-video, image-to-video, first-and-last-frame, subject reference, motion reference, voice reference, and video editing — with hard boundaries between image, video, and audio.

These task silos constrained creative workflows and capped the model's ability to generalize during training. MiniMax's first design principle for H3 was to unify and generalize across all of these tasks from the pretraining stage onward.

A Three-Module System

The H3 system is composed of three modules that work in sequence:

H3-Context-IR handles instruction parsing, cross-modal association, and temporal understanding. It takes free-form multimodal inputs — a prompt, reference images, video clips, audio clips — and distills them into a structured "Context Intermediate Representation." Source material requiring roughly 100,000 tokens of inference is compressed to about 4,000 tokens (a 25:1 ratio). Language acts as the generalizable bridge between context and target video. This module is hosted as a managed API rather than open-sourced.

H3-Base is the open-weight core: a 33.1-billion-parameter dense single-stream transformer that generates video and stereo audio at 768p. It uses Qwen3-VL-32B as its text and visual encoder (extracting hidden states from layer 50), a custom H3-VisualVAE for video tokenization, and an H3-AudioVAE for audio. Three-dimensional Multimodal Rotary Position Embeddings (MM-RoPE) handle positional relationships across temporal and spatial dimensions.

H3-Regenerate-2K takes the 768p output and the original multimodal context and regenerates a 2K version. Rather than using a conventional super-resolution module that guesses at missing detail, the base model regenerates in-context — drawing on the original reference material to recover fine textures, text, and brand marks that standard upscaling cannot reconstruct.

What Makes the Architecture Interesting

A few design choices in H3 are worth examining closely.

The VAE overhaul. MiniMax completely rebuilt its tokenizer for H3. The new H3-VAE achieves a 32× spatial compression factor and a 4× temporal compression factor, yielding a 4× gain in effective sequence length compared to the previous generation. This is what makes native 2K output economically viable — the sequence length stays manageable even at high resolution.

AdaLN precomputation. About 13 billion of the 33.1 billion parameters reside in Adaptive Layer Normalization (AdaLN) modulation branches. These can be precomputed and cached from the timestep embedding, bringing the effective inference parameter count closer to 20 billion — a meaningful efficiency gain.

Separated compute workloads. Introducing multimodal context tripled the variance in sequence length across training samples. H3 separates the compute workloads for understanding and generation, fine-tuning hardware utilization for each while balancing per-sample heterogeneous compute against load balancing across samples. MiniMax reports this lifted training throughput by nearly 30%.

Two checkpoint variants. The open-weight release on Hugging Face ships two task-specific checkpoints: H3-Base-FL2VA for first-and-last-frame generation (zero, one, or two input images), and H3-Base-Ref2VA for omni-reference generation (up to 9 images, 3 video clips, and 3 audio clips). The split keeps each checkpoint focused while sharing the same underlying architecture.

Input Specifications and Practical Limits

H3 accepts up to 9 reference images (JPG, PNG, WEBP, HEIC, up to 30 MB each), up to 3 video clips (MP4 or MOV, 2–15 seconds each, up to 50 MB each), and up to 3 audio clips (WAV or MP3, 2–15 seconds each, up to 15 MB each). Total files per request are capped at 12, with a 64 MB body limit and a 7,000-character prompt limit.

One constraint worth noting: audio cannot be sent as a standalone input — it must accompany at least one image or video clip. Image-to-video and reference-to-video modes are mutually exclusive within a single request.

Output is 4–15 seconds at 24 FPS with native 32 kHz stereo audio, in aspect ratios from 21:9 to 9:16.

Pricing and Access

The API uses an asynchronous task model: you submit a generation request, receive a task ID, poll for completion, and download the result URL. Pricing is $0.13 per second of 2K video output (roughly $7.80 per minute), with the first five reference images free and additional images at $0.04 each. Reference audio is free. At 768p, pricing drops to $0.09 per second, though that tier is currently in closed beta.

For comparison, Kling 3.0 at 1080p runs about $20.16 per minute and Dreamina Seedance 2.0 at 1080p runs about $22.45 per minute. H3's 2K output at $7.80 per minute is substantially cheaper, which is part of why it has attracted attention from commercial content teams.

Open Weights and License Considerations

MiniMax open-sourced the H3-Base weights on August 3, 2026, under the MiniMax H3 Community License. The license permits commercial use for organizations with annual revenue under $20 million, with prominent attribution required. Organizations above that threshold need a separate commercial agreement.

The H3-Context-IR and H3-Regenerate-2K modules remain hosted as managed APIs. Local deployment of H3-Base is supported via SGLang with four GPUs using Ulysses-degree parallelism, but the full 2K workflow requires API calls to the closed modules.

What This Means for Practitioners

H3's most practical contribution is collapsing the multi-model pipeline into a single context. A prompt like "Reference the camera movement from Video 1, have the character in Image 2 sing, with vocals matching Audio 3" is handled by one model rather than a chain of specialized systems — simplifying both engineering and the creative iteration loop.

The in-context regeneration approach to 2K output is also worth watching. Using the base model's own generative capability to recover detail rather than a separate upscaler could influence how other video generation systems handle resolution scaling.

The open-weight release, even with its revenue cap, is notable. Closed-source models have dominated video generation since the field emerged. H3 is the first commercially viable flagship to ship weights, opening the door to fine-tuning, hardware-specific optimization, and community experimentation that closed APIs cannot support.

Conclusion

MiniMax H3 is a technically coherent attempt to replace a fragmented ecosystem of specialist video models with a single unified system. The architecture choices — the rebuilt VAE, the AdaLN precomputation, the separated compute workloads, the in-context 2K regeneration — are each aimed at making that unification practical rather than conceptually appealing. Whether the approach holds up as the model scales and the technical report arrives will be worth following.

Weights are available at MiniMaxAI/MiniMax-H3 on Hugging Face, and the API is live at platform.minimax.io.