MiniMax H3 · Local
Weights are downloadable and ComfyUI has native support, so a Hailuo H3 local install is mostly just files in folders. What catches people is not the diffusion model — it’s the encoder sitting next to it. If you only need the published limits and a hosted-vs-local comparison before committing, the MiniMax H3 spec overview covers that in one page.
Size comparisons almost always quote the diffusion weights alone. But H3's text encoder is Qwen3-VL 32B, and it has to be resident while your prompt is encoded. That single file is frequently larger than the generation model people sized their card against — and it is why a machine that 'should' fit H3 on paper runs out of memory in practice.
If you are tight on memory, the INT8 or NVFP4 encoder builds buy back more headroom than dropping the diffusion model another quant level, because the encoder is the larger of the two.
The whole install is five steps and needs no custom nodes. This is the official ComfyUI procedure — the file table further down covers exactly what to download in step two, with sizes and official links.
Native H3 support landed on August 3, 2026. On older builds the H3 nodes simply do not exist, and the first error you hit will not tell you that — see the troubleshooting table below.
For text-to-video and image-to-video you need four files: the fl2va pruned INT8 diffusion model, the NVFP4 text encoder, and both VAEs. Add the ref2va checkpoint only if you want reference-driven generation. Every file links to its official Hugging Face page in the table below, with its exact size.
diffusion_models, text_encoders and vae — the repack is packaged so nothing needs renaming. The folder map below shows exactly which file goes where.
Restart ComfyUI, then open Template Library → Video and pick "MiniMax H3 T2V", "MiniMax H3 I2V" or "MiniMax H3 R2V". The templates ship with ComfyUI itself; if one shows missing-model warnings, it names the exact file it expects.
The stock template settings (768px short edge, around 5 seconds) are the configuration the tutorial documents. Console messages about dtype fallbacks are expected per the official tutorial and do not affect output. Get one clean run first, then change one variable at a time.
The official tutorial says Sage Attention roughly doubles generation speed with minimal quality loss. Two ways to enable it:
MiniMax’s official H3 model page recommends ComfyUI, SGLang, vLLM and Diffusers as local deployment paths. The table distinguishes their documented interfaces; Lightloom has not benchmarked their speed, memory use or output quality.
| Framework | Interface | What MiniMax links |
|---|---|---|
| ComfyUI | Visual node graphs and MiniMax templates | MiniMax links the ComfyUI tutorial and templates; the file map below follows that packaging. |
| SGLang | Serving-oriented local deployment | MiniMax links a dedicated H3 cookbook, including its documented hybrid workflow. |
| vLLM | Recipe-driven inference from original checkpoints | MiniMax points to the H3 recipes maintained by the vLLM project. |
| Diffusers | Python pipeline using the Diffusers-format release | MiniMax lists a Diffusers-format path in the official H3 model card; the linked implementation remains an upstream integration surface. |
First-party source: MiniMax H3 local deployment page
The official repack is packaged to drop straight into ComfyUI’s model folders. You do not need every file in the first group — pick the variants that match your workflow and your memory budget. Every filename links to its official Hugging Face page, and sizes are the exact ones reported by the Hugging Face file API, in decimal GB.
The generation model itself. Pick ONE fl2va file and/or ONE ref2va file depending on which workflows you plan to run. The repo README says to prefer int8_convrot variants when your PyTorch build supports cu130, and to fall back to fp8_scaled only when it does not.
Required. This is a 32B model — see the note below, it is the single biggest reason people underestimate what H3 needs.
Both required. H3 generates picture and sound together, so it needs a VAE for each.
Optional. Turbo LoRAs cut denoising steps once the base install works — none of them is needed for a first run.
Source: Comfy-Org/MiniMax-H3 (official ComfyUI repack) · requires ComfyUI 0.30.0 or later.
The set below is the one ComfyUI's own tutorial specifies — the pruned INT8 diffusion model and the NVFP4 text encoder, not the bf16 files. It is the smallest official combination that runs the T2V and I2V workflows.
Add the ref2va diffusion model on top if you want reference-to-video — it is a second checkpoint of comparable size, not a small add-on.
Source: File set per docs.comfy.org · sizes per the Hugging Face file API
The Comfy-Org repack uses the same directory names as ComfyUI's model tree, so the Hugging Face CLI can drop every file straight into place. Unlike a browser download, hf download resumes a dropped connection instead of restarting a 21GB file from zero.
hf download Comfy-Org/MiniMax-H3 \
--include "diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors" \
"text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors" \
"vae/*" \
--local-dir ComfyUI/modelsThat is the tutorial's 42.48 GB working set, exactly. Run it from the directory that contains ComfyUI (or swap in an absolute path), and swap the include patterns if you picked different variants above. Add loras/* only after the base install runs.
Every guide tells you which folder each file goes in; almost none tells you what you agree to by downloading them. The weights ship under the MiniMax H3 Community License Agreement — not Apache or MIT — and four of its terms are directly relevant to a MiniMax H3 local deployment:
This is a summary of the published license text, not legal advice. The full agreement is a short read — check it before building anything commercial on these weights.
Sources: MiniMax H3 Community License Agreement · official model card
These are the failures people actually report in the first hour of a MiniMax H3 local install — with the cause the error text does not state. Every entry links to its source, GitHub issues first, so you can check the current status yourself. Lightloom has not reproduced these on its own hardware.
ComfyUI is older than 0.30.0 — native H3 nodes are not present in earlier builds.
FixUpdate ComfyUI, restart, and reload the template.
Source: docs.comfy.org tutorial
The workflow JSON is newer than your ComfyUI core. The error text never hints at this, which is why it sends people hunting for corrupt downloads.
FixUpdate ComfyUI — this is a version mismatch, not a broken file.
Source: ComfyUI issue #15378
The audio VAE is missing or not connected. H3 decodes picture and sound separately, so it needs both VAE files.
FixInstall minimax_h3_audio_vae_fp32.safetensors and check the audio VAE output actually feeds the video-save node.
Source: ComfyUI Wiki guide
An fl2va checkpoint is loaded where the reference workflow expects ref2va weights.
FixDownload the ref2va diffusion model — it is a separate checkpoint, not a mode switch on the one you have.
Source: kingy.ai H3 guide
VRAM and system RAM are both constraints, and video references spike VAE memory beyond what the file sizes suggest.
FixChange one variable at a time: a smaller quant file, then lower resolution or duration, then flags like --lowvram or --disable-pinned-memory.
Source: ComfyUI issue #15312
H3's documented floor is 384p. At 256p runs do not degrade — they fail.
FixKeep the short edge at 384px or above.
Source: ComfyUI Wiki guide
A reported platform issue on the MPS backend. The same setups run I2V and R2V without it.
FixNo documented fix at the time of writing — follow the issue for status before buying disk space for the download.
Source: ComfyUI issue #15315
Sources: MiniMax H3 repo README · ComfyUI issue #15315
The open checkpoints are H3-Base, which generates at a default short edge of 768px. The 2K clips in MiniMax's launch material go through H3-Regenerate-2K, a separate module the official README lists as not yet open-sourced — the same is true of the H3-Context-IR understanding module. No MiniMax H3 local install produces native 2K today; pipelines that reach 2K route back through MiniMax's hosted API.
Source: MiniMax-AI/MiniMax-H3 README
A MiniMax H3 local install is worth it if you generate constantly and 768p output is enough. When a clip needs 2K, the render routes through hosted infrastructure either way — Lightloom’s generator runs H3 with 2K output and native audio, with none of the download or maintenance above.
Compare hosted pricing →Once the files are in place, the ComfyUI workflow guide covers the three templates and the errors that show up first. If memory is your constraint, the VRAM breakdown explains what actually allocates during generation, and the MiniMax H3 LoRA guide documents two current v4 paths that reduce denoising from roughly 20 steps to a 6–8-step starting range: Larry’s original custom-node workflow and drbaph’s specific pruned conversion. That does not prove compatibility with every checkpoint above, and total wall-clock speed still depends on model loading, decoding and audio work.
With the install done, the thing that decides whether your first clip is any good is the prompt, not the checkpoint. The MiniMax H3 prompt guide separates the three-field base format from full-reference R2VA’s six-section contract, with a library for both — including the audio blocks that stop H3 inventing a soundtrack for you.
Four things: one diffusion model (fl2va for frame-driven, ref2va for reference-driven), the Qwen3-VL 32B text encoder, the video VAE and the audio VAE. The encoder and both VAEs are non-optional — the diffusion model alone will not run.
Diffusion models go in ComfyUI/models/diffusion_models/, the text encoder in ComfyUI/models/text_encoders/, and both VAE files in ComfyUI/models/vae/. The Comfy-Org repack is packaged specifically to drop into those three folders.
Because the text encoder is Qwen3-VL 32B. Most size comparisons quote only the diffusion weights, but the encoder must be resident while your prompt is encoded and is frequently the larger of the two files. That is why machines that look sufficient on paper still run out of memory.
No. The open checkpoints are H3-Base, which outputs at a default short edge of 768px. The 2K clips in MiniMax's launch material go through H3-Regenerate-2K, a separate module the official README lists as not yet open-sourced — so 2K currently requires a hosted pipeline, not a bigger GPU.
Downloading the weights means accepting the MiniMax H3 Community License Agreement. It permits commercial use up to 20 million US dollars in yearly revenue, but its applicable territory excludes the United States, the European Union, the United Kingdom and South Korea — the official model card links a separate application form for those four regions. The outputs you generate are yours; MiniMax claims no rights over them.
Partially, and largely undocumented. On Apple Silicon the MPS backend runs the I2V and R2V workflows, but T2V has an open ComfyUI issue producing black video with NaN audio and no documented fix. For AMD there is no official H3 path at all — every first-party install document assumes an NVIDIA GPU with CUDA.
Pruned INT8 checkpoints are roughly 40% smaller than standard INT8 thanks to precomputed adaLN curve tables. They are a good way to save space, and they do not lock you out of the turbo LoRA — the publisher supports pruned and curve variants alongside the full ones, with the ComfyUI node detecting a pruned base automatically.