MiniMax H3 in ComfyUI: Open Weights, Requirements and Limits
MiniMax H3 works in ComfyUI natively — support was merged the same day the open weights landed, on August 3, 2026. But "open source" here has a specific shape: the H3-Base model is open under the MiniMax H3 Community License, while two pieces of the hosted pipeline are not released at all. This page explains what you actually get locally, what hardware it takes, and where the community builds come from.
Open weights, VRAM, GGUF, LoRA — the practical picture.
MiniMax H3, ComfyUI and Local Deployment
This is not a step-by-step install guide — dedicated ComfyUI documentation sites cover installation better than we can. What follows is the accurate status of each thing people search for before they start.
Does MiniMax H3 work in ComfyUI?
Yes, natively. MiniMax H3 launched on the MiniMax API and the Hailuo App on July 31, 2026 as a hosted-only model. On August 3, 2026 MiniMax released open weights for H3-Base, including the FL2VA and Ref2VA checkpoints, under the MiniMax H3 Community License — and native ComfyUI support was merged the same day via pull request. You do not need a third-party custom node pack to load the base model.
What is open, and what isn't
Open: the H3-Base model weights with the FL2VA and Ref2VA checkpoints, roughly 33B parameters, under the MiniMax H3 Community License. Not open: H3-Context-IR, the prompt preprocessing and enhancement system, and H3-Regenerate-2K, the 2K upscale module. Both remain hosted-only features. A local ComfyUI setup running the open weights therefore does not reproduce the full hosted pipeline — in particular, the 2K output path and the prompt-enhancement layer are not part of what you download.
How much VRAM does MiniMax H3 need?
MiniMax has not published official per-precision VRAM requirements, so every number circulating is community-reported and should be treated as an estimate rather than a spec. At roughly 33B parameters, full-precision inference is firmly in high-end-GPU territory. Below that, the community has produced INT8 and pruned INT4 builds, GGUF quantizations spanning Q2 through Q5, and NF4 4-bit builds — with NF4 reported to run on as little as around 8GB of VRAM. That 8GB figure comes from community reports, not from MiniMax, and lower quantization levels trade output quality and speed for the reduced footprint.
Is there an official MiniMax H3 GGUF?
No. MiniMax's open release covers H3-Base and its checkpoints — it does not include GGUF files. Every GGUF build of MiniMax H3 in circulation, along with the INT4/INT8 prunes and NF4 builds, was produced by the community after the August 3 release. That means quality, naming and correctness vary between uploads, and there is no vendor-blessed reference build to check against. If you go this route, weigh who published a given quantization and how it was produced before trusting it.
MiniMax H3 LoRA and H3 Turbo
MiniMax H3 Turbo is a distilled variant that generates in 4 steps, reported at roughly a 3.5x speedup over the standard model, with a corresponding Turbo LoRA available. Beyond that, the usual LoRA ecosystem dynamic applies: because the base weights are open, community-trained LoRAs appear over time for specific styles and subjects, with the same source-reliability caveat as the quantized builds.
Can I use MiniMax H3 without ComfyUI?
Yes — the model is also available as a hosted service through the MiniMax API and the Hailuo App, which is where the full pipeline including Context-IR and 2K upscale lives. If your goal is to generate video rather than to run models locally, a hosted platform removes the GPU, the downloads and the environment entirely. On Wowovid, MiniMax H3 text-to-video is live right now at 768P or 2K, 4 to 15 seconds, straight from the browser.
What Each Path Actually Costs You
Running open weights locally is the right choice for some work — full control, offline use, custom LoRA training. It is the wrong choice if you just want the clip. Here is the honest split.
Running MiniMax H3 in ComfyUI
- A capable GPU with enough VRAM for your chosen precision — full precision needs high-end hardware, and lower-VRAM options mean community quantized builds with quality trade-offs.
- Tens of gigabytes of model downloads for a ~33B-parameter model, plus checkpoints and any LoRAs.
- A working ComfyUI installation, correct node versions, and a workflow graph you maintain yourself.
- No H3-Context-IR prompt enhancement and no H3-Regenerate-2K upscale — those are hosted-only and not part of the open release.
- Troubleshooting is on you: driver versions, VRAM errors, and community builds of varying provenance.
Generating on Wowovid
- Open the page in a browser. No GPU, no drivers, no VRAM budget to plan around.
- Nothing to download and nothing to keep updated — no checkpoints, no node packs, no quantized builds to vet.
- MiniMax H3 text-to-video is live: write a prompt, pick 768P or 2K and a length from 4 to 15 seconds, and the result is saved to your account.
- One credit balance across every video and image model on the platform, not a per-model setup.
- HD export without a watermark on paid plans.
MiniMax H3 Local Deployment — FAQ
The questions that come up most when people research running MiniMax H3 on their own hardware.
MiniMax has not published official VRAM requirements, so treat every figure you find as a community estimate. MiniMax H3 is roughly a 33B-parameter model, which puts full-precision inference on high-end GPUs. Community-produced INT8, pruned INT4, GGUF Q2–Q5 and NF4 builds lower the requirement substantially — NF4 4-bit has been reported to run on around 8GB of VRAM — at the cost of output quality. Because these builds are community-made rather than official, expect variation between them.
Non-NVIDIA setups are consistently harder for large video models: the tooling around ComfyUI, quantization formats and attention kernels is built primarily against CUDA, and Apple Silicon and AMD paths depend on community support that lags behind and varies in maturity. MiniMax has not made platform-support claims beyond the open weight release itself, so we won't state a definitive answer here. If your hardware is not an NVIDIA GPU with substantial VRAM, generating through a hosted platform will usually be the shorter path.
MiniMax released the open weights for H3-Base on August 3, 2026, including the FL2VA and Ref2VA checkpoints, under the MiniMax H3 Community License — distributed through MiniMax's official channels. Always start from MiniMax's own repository or model page rather than a mirror: the quantized and GGUF versions circulating are community re-uploads, not official releases, and their quality and provenance differ. Note again that H3-Context-IR and H3-Regenerate-2K are not included in any download — they exist only in the hosted service.
MiniMax H3 is approximately 33B parameters. On disk that means tens of gigabytes at full precision, which is why the community moved quickly to quantized formats — GGUF at Q2 through Q5, INT8 and pruned INT4, and NF4 — each shrinking the download and the VRAM footprint while giving up some output quality.
MiniMax H3 Turbo is a 4-step distilled variant of the model, reported at roughly a 3.5x speedup over the standard generation path, with a matching Turbo LoRA. It is aimed at cutting generation time, which is the main pain point of running a model this size locally.
Yes. MiniMax is offering the model through its own API and the Hailuo App, and hosted platforms let you generate video in a browser with no local setup at all. On Wowovid, MiniMax H3 text-to-video is available right now — no GPU, no downloads and no ComfyUI environment — at 768P or 2K and 4 to 15 seconds per clip. See our MiniMax H3 page for what the model does and how it compares to Kling AI and Seedance 2.5.
Independent Platform
"MiniMax H3", "MiniMax" and "Hailuo" are trademarks of MiniMax and are used here for descriptive purposes only. Wowovid is an independent platform offering dozens of image and video generation models, and is not affiliated with, endorsed by, or sponsored by MiniMax. ComfyUI is an independent open-source project, likewise unaffiliated with Wowovid.
Skip the Setup
If the goal is the video and not the environment, generate in your browser instead — MiniMax H3 text-to-video is live on Wowovid, with no GPU, no downloads and no workflow graph to debug.