MiniMax H3 · Model overview
MiniMax H3, released as Hailuo 3.0 on July 31, 2026, is the model behind the Lightloom generator. This page collects what it actually is, its published limits, what running it yourself really takes, and the five guides that cover the parts the release notes skip.
MiniMax H3 is an omni-modal model: text, images, video and audio share one context, and picture and sound come out of a single forward pass rather than being stitched together afterwards. That is why the audio lines up with the action — and why leaving the audio block out of your prompt doesn’t give you silence, it gives you a soundtrack the model picked.
It is also why MiniMax H3 is heavier to run than a video-only model of the same nominal size. The audio branch is not free, and neither is the 9-image reference path that makes the third generation mode work.
Published limits, not measurements — this site has no first-party benchmark hardware, so nothing here is a timing claim.
| Maximum resolution | 2K |
|---|---|
| Frame rate | 24fps |
| Maximum duration | 15 seconds per generation |
| Resolution cap | 768px short edge, capped at 768×1344 |
| Audio | Native stereo, generated with the picture |
| Reference inputs | 9 images, 3 videos, 3 audio clips |
| Released | July 31, 2026 |
Source: docs.comfy.org / MiniMax release notes.
MiniMax H3 weights are downloadable, so running it yourself is a real option — it is just a much larger commitment than the download page suggests.
| Lightloom hosted T2V | MiniMax H3 on your GPU | |
|---|---|---|
| Hardware | Remote MiniMax H3 API execution; no local GPU required. | MiniMax's own configuration calls for 4 GPUs. Community GGUF builds start at 8.49GB of weights before overhead. |
| Setup | Google sign-in and a paid credit balance; no model download. | ComfyUI 0.30.0 or later, plus four model files placed across three folders. |
| Output | Text-to-video at 768P or 2K, 4–15 seconds, with native audio. | Same model, same ceiling — if your card can hold it at that resolution. |
| Cost | 20 credits/second at 768P or 30 credits/second at 2K. No subscription and no free allowance. | No per-clip cost, once you own hardware that can run it. |
| Control | Text-to-video is available; image and multimodal reference inputs remain guide-only until secure uploads ship. | Full node graph: samplers, sigma shift, LoRAs, quantization. |
Community GGUF builds of MiniMax H3 go as small as 8.49GB, and every roundup on the internet quotes that number as the requirement. It is not one. A weight file is not a VRAM figure — the VAE, text encoder, latents and the audio branch all allocate on top of it, and peak usage climbs with resolution and duration. The MiniMax H3 VRAM breakdown has every quant level against the card it actually fits, with the overhead nobody lists. If you do have the hardware, the MiniMax H3 ComfyUI workflow guide walks through all three official templates and what breaks on older versions.
Two things decide whether a local install is pleasant or painful. The first is that the text encoder is a 32B model — usually larger than the generation model people sized their card against, which the MiniMax H3 local setup guide covers file by file. The second is sampling time: the MiniMax H3 turbo LoRA cuts the roughly 20-step base workflow to a current 6–8-step starting range. Four steps remains a documented minimum, where audio artifacts are more likely and quality needs closer checking.
Lightloom sends an allowed text prompt to the official MiniMax H3 V2 API, so local VRAM and ComfyUI versions are not part of the user workflow. The service reserves credits before submission, stores completed clips in its own Cloudflare bucket and returns the reservation if the task fails. Lightloom is independent and is not affiliated with MiniMax.
Prompting is the one part you still own either way — the MiniMax H3 prompt guide documents the official three-field format with copy-paste examples that load straight into the generator.
Running MiniMax H3 on your own hardware is where most people get stuck. These cover the parts the release notes skip — what actually allocates VRAM, how low-step Turbo settings affect audio, and which of the 15 model files you really need.
MiniMax H3 (also called Hailuo 3.0) is an omni-modal generation model released on July 31, 2026. It understands text, images, video and audio in a single context and generates video with native stereo sound at up to 2K resolution, 24fps, and about 15 seconds per clip.
Yes. Hailuo 3.0 is the product name MiniMax uses for the same release; technical documentation and community tooling mostly say MiniMax H3, while the official app and announcement use Hailuo. Specs, license and API refer to one model.
Up to 2K at 24fps, with the short edge capped at 768px and an overall cap of 768×1344. Requests above that get clamped back down, so pick an aspect ratio inside the cap and upscale afterwards if you need more.