Short answer: MiniMax’s hosted H3 service moderates requests. Running open weights in a fully local workflow changes where moderation can occur, but the published documents do not establish a universal “uncensored” capability or guarantee that a particular prompt will generate. Choose hosted access for a managed workflow; investigate local deployment only after checking the exact components, hardware and applicable release terms.
Compare more models: Can You Run It at Home? lists every Kingy verdict by machine type, from a 16GB laptop to a 512GB Mac Studio.
Evidence boundary: No Kingy hands-on test or generated sample is presented in this article. The hardware rows in the box above are linked third-party reports, not Kingy measurements. The moderation analysis uses MiniMax’s published files and licensing terms, ComfyUI documentation, and clearly labelled community reports. Those reports are anecdotal rather than a controlled benchmark.
Start with the deployment you actually intend to use. MiniMax’s API, Comfy Cloud, a local ComfyUI template and the raw official checkpoints have different infrastructure and terms. A local-looking interface can still call hosted services. The table below separates the documented routes.
Check the route behind the interface
Identify which parts of your H3 workflow run locally and which call a hosted service.
Inspect the local H3 setup and its limits
This checklist adds no generated sample or performance test. The article’s hardware reports remain third-party evidence.
How to inspect the workflow route
For an H3 workflow, record where each stage runs before relying on a local or hosted label. A desktop interface can include network-backed nodes, and a local model’s documented behavior does not predict a hosted service’s moderation result.
- Name the generator, any prompt-processing step, and any upscaler. Identify which use local files and which call a service.
- Check the exact release terms and the handling of your reference media. Use a non-sensitive test clip for an unfamiliar route.
- Plan a short, ordinary shot you can evaluate for motion, continuity and audio. Record refusals separately from model-load or workflow errors.
The current ComfyUI H3 documentation is the starting point for current templates. This article’s hardware reports remain third-party evidence, and this checklist adds no generated sample or performance measurement.
Next: Choose a generator · H3 setup
Which H3 route answers your question?
| Route | What the documents establish | Useful next step |
|---|---|---|
| MiniMax hosted H3 API | Managed requests with moderation and documented API limits; 768P or 2K output; MiniMax does not disclose the precision it serves | Check the model/task fields and retain the returned status; a refusal differs from a technical error |
| ComfyUI open-weight template | Native 768px-short-edge generation; named checkpoint families and version requirements | Use the current template’s component list and inspect whether any node calls a hosted service |
| 2K delivery | MiniMax’s hosted H3-Regenerate-2K stage, which regenerates the 768p base result at 2K, or a separate third-party upscale pass on local 768p output | Keep base generation and enhancement steps separate when reporting privacy, cost and resolution |
What MiniMax H3 actually is
MiniMax H3 is a multimodal video-generation system that combines text, images, video, and audio. The official model supports text-to-video, first-frame and last-frame image-to-video, and reference-driven generation using images, video clips, and audio. It generates video with synchronized stereo audio rather than requiring a separate audio pass.
The documented output range is four to fifteen seconds at 24 frames per second. The local H3-Base checkpoints (FL2VA and Ref2VA, published by MiniMax in BF16) are designed around 768p output. MiniMax’s full workflow reaches 2K with two hosted stages: H3-Context-IR prepares the prompt and references before generation, and H3-Regenerate-2K regenerates the 768p result at 2K. Neither stage is included in the local release; MiniMax offers both only through its API. The official H3 model repository and MiniMax API guide describe the current architecture and limits.
That makes H3 different from a lightweight image-to-video plugin. It is a large, multimodal generation system with separate task families for first/last-frame generation and reference-to-video generation. It can carry a subject, visual style, motion pattern, camera move, or voice across a short shot when the reference material and prompt are well matched.
Why the hosted version is not uncensored
The hosted H3 API uses a multimodal request structure: a text prompt can be submitted alone, or paired with first-frame, last-frame, reference-image, reference-video, or reference-audio inputs. The API then creates an asynchronous generation task and returns a downloadable result if the task succeeds.
That infrastructure also gives MiniMax an opportunity to moderate the request. The H3 documentation states that submitted text, images, videos, and enhanced prompts are subject to automated moderation. It specifically warns that content suspected of being unlawful, pornographic, or infringing third-party rights may be blocked. The API documentation includes a sensitive-content error response for rejected video descriptions.
In other words, changing the wording of a prompt is not the same as removing moderation. The hosted service remains a controlled endpoint. It may block a prompt before generation, reject an uploaded reference, or apply additional review to the result.
The hosted route is still useful. It is easier to start, is the only official route to 2K output (MiniMax’s hosted H3-Regenerate-2K stage; the API also serves 768P), and avoids the hardware burden of downloading and running the checkpoints. But it is not the route to choose if the specific research question is whether H3 behaves differently without MiniMax’s server-side filter.
What changes when H3 runs locally
Local inference changes the control boundary. Instead of sending the prompt and reference files to MiniMax’s generation infrastructure, you download the model components and run them through a local interface such as ComfyUI, Diffusers, SGLang, or vLLM.
That means there is no MiniMax API request to inspect at generation time. The local workflow can still fail for ordinary technical reasons—unsupported input, insufficient memory, bad dimensions, a malformed workflow, weak prompt adherence, or a model limitation—but the failure is no longer necessarily a hosted safety refusal.
This is why early community reports describe local H3 as unusually permissive. Several public forum threads report successful generations that users say would have been rejected by more heavily moderated hosted video tools. Those observations are worth recording, but they are not a controlled benchmark. They may reflect different checkpoints, wrappers, quantizations, prompts, seeds, or the absence of a platform-level filter.
Community reports can suggest test cases, but they do not establish how this checkpoint behaves across prompts, seeds or wrappers. This article makes no comparative “most permissive” ranking. The repository also restricts bypassing safeguards; removing a hosted API call does not remove the release terms.
Beginner workflow: run H3 through ComfyUI
ComfyUI is currently the most approachable local route because it exposes H3 through ready-made workflow templates. The ComfyUI H3 guide documents native text-to-video, image-to-video, and reference-to-video workflows.
1. Check the license before downloading
Read the MiniMax H3 Community License Agreement first. The current license applies worldwide except for the United States, European Union, United Kingdom, and South Korea. Those excluded territories require separate authorization from MiniMax. Canada is not listed among the excluded territories in the current agreement, but users should verify the current license and local requirements before deploying the model.
This is not a minor footnote. Open weights are not automatically equivalent to an unrestricted open-source license. The H3 agreement contains territorial limitations, use restrictions, commercial terms, redistribution conditions, and obligations for services that expose the model to other users.
2. Install a current ComfyUI build
ComfyUI’s current guide specifies 0.30.0 or later for its base Text to Video, Image to Video and Reference to Video templates; Multiframe Reference needs 0.34.0, newer control/attention nodes need 0.35.0, and FastH3 needs 0.36.0. Match the template and required nodes to your build. Current ComfyUI H3 guide
The first launch may download several large model components. H3 is not a small desktop model, so storage, system RAM, GPU memory, and generation time matter. Start with the template’s 0.4-megapixel preview (864×480 at 16:9) before attempting the native 1344×768 canvas.
3. Download the official model components
The raw MiniMax repository provides FL2VA and Ref2VA task families in BF16. ComfyUI’s templates use their own component distribution in Comfy-Org/MiniMax-H3: the templates load pruned int8-convrot builds, compressed from the BF16 release, and the same repository also lists full BF16 builds. Use the template’s actual component list; the raw-repository command below is an alternative download layout, not a drop-in ComfyUI installation command.
hf download MiniMaxAI/MiniMax-H3 \
--include "model_index.json" \
"FL2VA/*" "Ref2VA/*" \
--local-dir MiniMax-H3Choose the family your task needs before downloading. In ComfyUI, keep checkpoint, LoRA and attention backend compatible: the October 26 guide warned that full-build LoRA adaln weights do not merge into pruned builds, and that Comfy Kitchen attention fails with the template’s int8-convrot checkpoints. Current compatibility of those combinations has not been reverified after the guide was reorganized. Check the template’s current component requirements before choosing a backend; a shape or alignment error does not establish a refusal.
4. Start with text-to-video
For a first test, use a harmless, well-defined scene. Describe the location, subject, action, camera movement, timing, and sound in one prompt. H3’s prompting guidance recommends describing the overall scene first and then breaking the action into timed shots.
For example:
A five-second documentary shot in a rain-soaked mountain observatory at blue hour. The camera begins in a wide exterior view, slowly pushes through the doorway, and settles on a scientist adjusting a telescope. Wind moves the rain across the glass. Natural room tone, distant thunder, soft mechanical clicks. Realistic lighting and restrained camera motion.This kind of test establishes whether the workflow works before you investigate moderation behavior. It also produces a useful baseline for comparing local H3 with the hosted API.
5. Try image-to-video and reference-to-video
Image-to-video is useful when you want to control the opening frame or the transition between a first and last frame. Reference-to-video is useful when you want to carry a fictional character, a visual style, a motion pattern, or a voice into a new shot.
Only use reference material you have the right to use. A local workflow makes it technically easy to upload a person’s photo or a commercial clip, but it does not grant permission to reproduce someone’s likeness, copyrighted work, or private material.
6. Understand the local resolution tradeoff
ComfyUI recommends a 768px short edge and a 1344×768 canvas for native 16:9 generation, close to H3’s trained resolution. That is a quality target, rather than a hard output-size limit: the H3 nodes accept width and height in 32-pixel steps up to 16384 per axis. The guide captured October 26 warned that larger sizes render more slowly without adding detail and can reduce quality; Kingy has not tested those larger sizes. Reference videos follow separate resizing rules based on a 768px short edge and 768×1344 target area, rounded to the 32-pixel grid, without upscaling smaller inputs. For 2K delivery, start at native size and upscale separately, or use MiniMax’s official hosted pipeline with H3-Context-IR and H3-Regenerate-2K. A separate upscaler does not establish native 2K detail from the base model. ComfyUI overview. Native workflow target; node schema and reference scaling. The 16384-pixel upper bound is an input setting, not a verified usable output size.
That creates an important tradeoff:
- Local base inference gives you more control and keeps the primary generation step on your machine.
- MiniMax’s hosted route adds H3-Context-IR, which its model card calls critical to output quality, and H3-Regenerate-2K for 2K output. MiniMax does not disclose the precision it serves, and prompts and media pass through its infrastructure again.
- A completely local pipeline currently means accepting H3-Base’s 768p output, using an unofficial upscaler, or waiting for MiniMax to open-source H3-Regenerate-2K, which its model card says is not yet released.
Hardware and performance expectations
H3’s local release is best understood as a serious workstation workload rather than a one-click browser feature. The exact requirement depends on the task family, precision, quantization, resolution, and runtime, and the official documentation does not promise one universal VRAM number. A machine that can load the template’s pruned INT8 checkpoint at preview size may still struggle with a longer clip, reference inputs, or the native 1344×768 canvas.
Plan around three constraints: model storage, memory headroom, and patience. Keep enough free disk space for the checkpoints and cache files. Start with short preview generations at a modest resolution. Then increase duration or quality one variable at a time. If a workflow fails, first separate a memory or configuration problem from a content refusal. A crash, out-of-memory error, or malformed output is not evidence that H3 filtered the prompt.
This distinction matters in articles about “uncensored” models because local workflows often circulate as screenshots or isolated clips. A successful render proves that one configuration produced one result. It does not establish a universal content policy for every quantization, wrapper, GPU, or future checkpoint.
Does local H3 create adult or controversial video?
The exact generation boundary is unmeasured here. Neither a successful community clip nor the absence of an API call proves that every adult or controversial prompt works. Test a lawful, ordinary creative task first and report the actual checkpoint, workflow, seed, outcome and error category.
Video models do not have a single censorship switch. Restrictions can enter at several points:
- The prompt can be filtered by a hosted service.
- Uploaded images, videos, or audio can be classified before generation.
- The model may have learned refusals or content aversions during training.
- A wrapper such as ComfyUI, a cloud provider, or a third-party interface may add its own filters.
- The generated output may be inspected after inference.
Local H3 primarily removes the first and, depending on the workflow, some of the second and fifth layers. It does not guarantee that the model will follow every prompt. It also does not make a prohibited use lawful.
For an evidence-led review, test categories rather than chasing a single viral example. Use several seeds and record whether the local and hosted routes produce, reject, or distort the same benign mature-theme prompts. Keep the test set legal and consensual. Do not test with minors, non-consensual sexual material, private images, or real-person impersonation.
Privacy: local does not mean risk-free
Local generation can improve privacy because the main prompt and reference files do not need to be uploaded to MiniMax. That is valuable for private concept work, unreleased product footage, or sensitive creative material.
But local privacy depends on the complete setup. A ComfyUI extension may download models or call an external service. A hosted upscaler may receive the output. A cloud GPU instance may retain logs or disks. A shared machine may expose the input and output directories. Users should check the workflow, extensions, runtime logs, network connections, and storage policy before assuming that a local interface is fully offline.
For sensitive work, use a clean environment, minimize extensions, disable unnecessary network access, and keep reference material in a controlled directory. Never upload a private person’s image merely because a workflow accepts it.
Consent, likeness, and deepfake risk
The H3 license prohibits impersonating another person without consent, authorization, or a lawful right to do so. That aligns with the basic rule creators should follow even where enforcement is unclear: do not use a real person’s face, voice, or private footage in a generated video without permission.
Reference-to-video makes this especially important because it can preserve a subject’s identity, motion, style, and voice across new scenes. The more convincing the output, the less defensible it is to treat the source person as a mere input asset.
Use fictional characters, licensed performers, consenting adults, or original assets. Label synthetic media clearly when it is published. Keep records of consent and source rights, especially for commercial work.
Copyright and commercial use
MiniMax’s license says that MiniMax claims no rights over outputs, but that does not guarantee that every output is free of third-party claims. The user remains responsible for the output and its use. A generated clip can still resemble a protected character, reproduce a recognizable logo, use an unauthorized likeness, or create a misleading imitation of a real advertisement.
Choose the license for your deployment: Comfy’s current licensing Q&A says the H3 Community License generally permits commercial use where its territorial and other conditions are met and annual commercial-product/service revenue does not exceed US$20 million. Local use in the United States, European Union, United Kingdom or South Korea needs a separate license; use of outputs is also subject to the agreement’s territory restrictions. The repository requires prior written authorization above that revenue threshold and a “MiniMax H3” interface notice for commercial products or services. Read the repository agreement for the conditions that apply to your deployment.
For teams needing a paid local license, Comfy lists Professional from US$5,000 per month, for up to ten licensed users and distilled open-weight versions. Enterprise pricing is quoted, with a twelve-month minimum. Comfy Cloud generations include commercial rights under that service’s terms. These are published licensing terms, not a license Kingy has purchased or a clearance for every intended use. Check your territory, revenue and distribution route against the current license page before committing to local deployment.
Those conditions are separate from any rights a user may have in a particular output. Before shipping a paid product, read the current agreement with counsel or the organization’s licensing team.
How to report the results honestly
If you publish a guide or review, make the experiment reproducible:
- Record the H3 repository revision and workflow version.
- Identify whether the test used local H3, the MiniMax API, Hailuo, or a third-party host, and whether the output is the 768p base result or a 2K regeneration.
- Record the task family: T2V, FL2VA, I2V, or Ref2VA.
- State the resolution, duration, seed, sampler or inference settings, and hardware.
- Run multiple seeds instead of presenting one successful clip as the norm.
- Separate refusals from technical failures and from visibly poor generations.
- Do not publish disallowed test material merely to prove a point.
Most importantly, label the evidence correctly. “The API rejected this prompt” is a verified observation. “The local checkpoint generated this test scene” is a reproducible test result. “H3 is uncensored” is a broad conclusion that requires much stronger evidence than either of those observations.
Hosted or local: the decision
Use hosted H3 when you want the managed service described by MiniMax. Use a local template when you have checked its hardware, component, network and licensing requirements and need control over that deployment. This review does not establish a measured local-versus-hosted refusal rate.
The published release terms, runtime and task family remain part of that decision. Begin with the same benign five-second brief, save the settings and classify the outcome as generated, refused, technically failed or unusable. Keep source specifications and observed results separate.
For a first check, retain the base output before enhancement and compare subject, action, timing, sound and reference adherence with your brief. The example in this guide is a reusable test prompt, not a generated Kingy sample.
For a separate published Hailuo/MiniMax H3 full test—not a local-hardware or moderation benchmark—see the MiniMax H3 video receipt.
Sources and further reading
- Official MiniMax H3 model repository
- MiniMax H3 Community License Agreement
- MiniMax H3 license Q&A
- ComfyUI MiniMax H3 workflows
- MiniMax H3 API reference
- MiniMax video-generation guide
- Example of an early community report — anecdotal, not an independent benchmark
Reviewed October 17, 2026: Rechecked the official model card, API moderation, release terms and current ComfyUI workflow. Added route/version distinctions and removed unsupported comparative-permissiveness wording. Community reports remain historical anecdotes. No new generation or moderation benchmark was performed.
October 26, 2026, native-size clarification: Distinguished ComfyUI’s recommended trained canvas, its node output limits and reference-video resizing. Historical clip reports retain their original dates and settings. No generation or compatibility test was run.
Licensing clarification, October 19, 2026: Replaced the broad paid-license wording with Comfy’s documented community-license eligibility and paid local-license routes. Existing hardware reports, test limits and release terms retain their original evidence and dates.
October 30, 2026, source qualification: Recovered node-size/reference-input support from official code after the overview changed. Marked removed compatibility and larger-size guidance by its retained source date. No generation or compatibility test was performed.
Trending on Kingy
Keep reading with the stories getting the most attention now.
The Kingy Brief
Get the free AI price sheet.
25 AI plans and models on one page, from official pricing pages, updated every month. Free when you join The Kingy Brief: AI changes, original tests and one practical thing to try, Fridays at 09:00 Vancouver time.
Free · Double opt-in · Unsubscribe anytime
Signup help and newsletter schedule
After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Your price sheet download is on the page that opens. Check Spam or Promotions if you cannot find it.
Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.