📖 Table of Contents
1. TL;DR — The Verdict
MiniMax H3 is the first open-weight video generation model from MiniMax (the Hailuo team), released today with Day-0 native support in ComfyUI. It accepts text, images, video, and audio as inputs and generates video with real stereo sound, up to 2K resolution, and up to 15 seconds per clip — all in a single pass.
Bottom line: This is the most significant open-weights video release of 2026 so far. Native stereo audio, multi-modal reference control, and a 66% memory reduction that makes 2K video generation runnable on an RTX 3060 — three things no previous open video model combined. If you've been waiting for a serious local alternative to Veo and Sora, this is it.
2. Model Overview: MiniMax's First Open-Weight Video Model
| Property | MiniMax H3 |
|---|---|
| Developer | MiniMax (Hailuo team) |
| Generation | 3rd-gen video model (after Hailuo 01 and Hailuo 02) |
| Weights | Open (first open-weight release from MiniMax) |
| Inputs | Text, image, video, audio |
| Output | Video up to 2K, up to 15 seconds |
| Audio | Native stereo, generated in the same pass |
| ComfyUI support | Day-0 (v0.30.0+) |
| Local hardware | RTX 3060-class GPU (42.5 GB footprint, quantized) |
H3 is the third-generation video model from the Hailuo team, and the first one MiniMax has released with open weights. That choice matters: Hailuo 01 and 02 were strong closed models, but H3 opens up the entire stack — weights on Hugging Face, official ComfyUI workflows, and a quantization pipeline tuned for consumer GPUs.
The headline capability is cross-modal generation: H3 takes images, audio, and video together and resolves them against a prompt that explains how they relate. Real production work rarely draws on a single modality, and H3 is explicitly built for that reality.
3. Capabilities: Five Tasks Collapsed into One Model
MiniMax markets H3 as replacing five separate tools with one model:
- Text-to-video — prompt only, no reference material needed.
- Image-to-video — bring a single image to life with motion and camera work.
- First-and-last-frame — control the opening frame, the closing frame, or both, and let the model fill in the middle.
- Reference-to-video — supply reference images, video, or audio and carry a subject, a motion, or a voice through the clip.
- Motion transfer — the most interesting one for graph work: a reference video supplies movement (a camera move, a performance, a cutting rhythm) while the subject and style come from elsewhere.
Why motion transfer matters: combined with in-place editing, it means you can iterate on a shot — swap the subject, keep the camera move, change the style — without regenerating from scratch. That's the workflow that makes video generation practical for real productions rather than one-off demos.
4. Native Audio: Stereo Sound Is a Property of the Model
This is the differentiator most reviewers will focus on. In H3, audio is generated with the video in the same pass, in stereo — it's a property of the model, not a post-process bolted on afterward.
Most video models either output silent video (leaving you to add music or foley in an editor) or generate audio as a separate second pass. H3 collapses those into one. For a 15-second clip, that means dialogue, ambience, and stereo placement are all coherent with the visuals — a massive time-saver for anything from product demos to short-form content.
5. ComfyUI Day-0: Workflow Setup
Day-0 support is rare and worth calling out: H3 workflows were available in ComfyUI the same morning the weights dropped. Setup is a three-step process:
- Update ComfyUI to the latest version (0.30.0 or newer), or use Comfy Cloud.
- Download the official workflows from the template library: T2V · I2V · R2V (reference-to-video).
- Follow the note in the workflow to download the models and save them in the correct model directory, then write your prompt, connect any frame or reference inputs, and run.
Weights are hosted at Comfy-Org/MiniMax-H3 on Hugging Face. The official ComfyUI integration means you get the full node graph — frame control, reference inputs, audio output nodes — without writing any custom code.
6. Running Locally: 3060 and the 66% Memory Cut
Here's the part that makes H3 genuinely different. Getting a 2K video model to run on consumer hardware took real ML engineering on ComfyUI's side:
- Modulation weight pruning: the model's modulation weights (~40% of total parameters) were pruned and replaced with a functionally equivalent lookup table, dramatically shrinking memory with no loss in output quality.
- int8 convrot quantization: the weights ship with an accurate, efficient int8 quantization scheme.
- Custom kernels: reduce peak VRAM usage during inference.
- Dynamic VRAM offloading: moves tensors in and out of VRAM as needed.
The combined result: total memory footprint reduced by 66%, from 123.6 GB in full precision down to 42.5 GB with the smallest model variants. With dynamic offloading, that's enough for a next-generation 2K video model to run locally on a GPU like the RTX 3060.
Reality check: "Runs on a 3060" means it's possible with aggressive quantization and offloading — expect slower generation than cloud GPUs, and note the 42.5 GB figure covers the smallest variants. If you want fast iteration, Comfy Cloud or a rented GPU will still beat local. But the fact that the ceiling moved from "impossible" to "possible at home" is the story.
7. H3 vs Veo 3 vs Sora 2: Open vs Closed
| Criterion | MiniMax H3 | Veo 3 (Google) | Sora 2 (OpenAI) |
|---|---|---|---|
| Weights | ✅ Open (HF) | ❌ Closed API | ❌ Closed API |
| Max resolution | 2K | 1080p (API tiers) | 1080p |
| Max length | 15 s | ~8 s per shot | ~20 s |
| Audio | ✅ Native stereo | ⚠️ Add-on | ⚠️ Add-on |
| Local run | ✅ RTX 3060-class | ❌ No | ❌ No |
| ComfyUI | ✅ Day-0 | ❌ No | ❌ No |
| Best for | Local pipelines, full control | Polish + ecosystem | Creative range |
The honest read: Veo 3 and Sora 2 still hold advantages in raw polish and creative range, and their APIs keep improving monthly. But H3 wins on the axis that matters for builders — open weights, native audio, and local execution. If your workflow is "generate → iterate in a node graph → ship," H3 is currently the only frontier-class option that lives entirely on your hardware.
8. FAQ
Q: Is MiniMax H3 really open source?
The weights are open and hosted on Hugging Face under MiniMax's release. The exact license terms should be checked on the model card, but unlike closed APIs there are no per-generation fees and no content restrictions from a platform gatekeeper — you run it on your own hardware.
Q: What GPU do I need?
ComfyUI's optimizations bring the smallest variants down to a ~42.5 GB footprint with int8 quantization and dynamic VRAM offloading, which is why RTX 3060-class cards (12 GB VRAM + system RAM offloading) can run it. Bigger GPUs mean faster iteration; 24 GB+ cards will be noticeably more comfortable.
Q: Does H3 really generate audio, or is it lip-sync guesswork?
Audio is generated natively in the same pass as the video — dialogue, ambience, and stereo placement are outputs of the model itself, not a separate text-to-speech or foley pass. That's the key architectural difference from models where audio is added afterward.
Q: How does this compare to Hailuo 01 / 02?
H3 is the third generation and the first with open weights. Compared to its closed predecessors, the big additions are native stereo audio, the multi-modal reference system, and the local-execution optimizations. If you used Hailuo 02 via API, H3 is a generational step up in both capability and control.
Q: Where can I try it?
- Weights on Hugging Face (Comfy-Org/MiniMax-H3)
- ComfyUI 0.30.0+ with the official T2V / I2V / R2V workflows
- Comfy Cloud for instant browser-based testing
- AI Video Generation 2026: Sora vs Runway vs Pika comparison — how the broader video model landscape stacks up