Can MiniMax H3 Run Privately On A Local GPU? The US License Catch
MiniMax H3-Base can generate video locally, but its full 2K workflow uses hosted services and the public license excludes the US, EU, UK, and South Korea.
Yes, MiniMax H3-Base can generate video and stereo audio locally after its weights are downloaded. But the complete H3 workflow is not fully local: MiniMax's Context-IR preprocessing and Regenerate-2K stage remain hosted. There is another major catch: the published community license excludes the United States, European Union, United Kingdom, and South Korea unless MiniMax separately authorizes use.
Watch The 30-Second Summary
That makes “Can MiniMax H3 run privately?” at least four questions: Can the model execute on your hardware? Do prompts and reference files stay on that machine? Does the workflow call a hosted service? And are you licensed to use the weights where you live?
What Is Confirmed
MiniMax released downloadable H3 weights, and ComfyUI added native workflows for them. The official MiniMax H3 model card describes two H3-Base variants:
- H3-Base-FL2VA supports text-to-video, first-frame, last-frame, and first-and-last-frame generation.
- H3-Base-Ref2VA accepts combinations of images, video, audio, and text as references.
H3-Base produces video at a default 768-pixel short edge, 24 frames per second, for 4 to 15 seconds, with native 32 kHz stereo audio. ComfyUI's H3 guide provides local text-to-video, image-to-video, and reference-to-video workflows.
The official model card is equally clear about what was not released as local weights. MiniMax describes H3 as a three-part system:
- H3-Context-IR interprets complex text, image, audio, and video references.
- H3-Base generates a 768p audio-video result.
- H3-Regenerate-2K uses the result and original context to regenerate at up to 2K.
MiniMax says Context-IR relies on multiple hosted models and services, so it is not included in the open release. Regenerate-2K is also offered through an API rather than included in the published local workflow. A machine running only the released H3-Base weights can be local; the official end-to-end 2K system is not.
The license boundary is also explicit. MiniMax's H3 Community License Agreement, dated August 2, 2026, defines the United States, European Union, United Kingdom, and South Korea as “Excluded Territories.” It says the agreement grants rights only in the remaining “Applicable Territory” and does not authorize use outside it.
That is not a minor attribution condition. If you are in an excluded territory, downloading the weights does not by itself give you permission under the published community license to run them. MiniMax's license Q&A says people and organizations in those regions can contact the company to apply for separate authorization.
What Is Still Unclear
Several practical questions do not have one universal answer.
- Minimum usable hardware: ComfyUI supports model offloading and publishes quantized weights, but its H3 guide does not promise one minimum VRAM figure, speed, or resolution for every workflow.
- Performance on a particular PC: generation time depends on GPU architecture, VRAM, system RAM, quantization, attention implementation, resolution, duration, operating system, and whether weights are repeatedly moved between RAM and VRAM.
- Privacy of third-party workflows: a JSON workflow, custom node, extension, model manager, telemetry package, or remote upscaler can add network calls that are not part of H3-Base itself.
- Hosted H3 retention for a particular route: MiniMax's Open Platform privacy policy describes collection of messages and content, service operation, security, analytics, sharing, and purpose-based retention. It does not provide one simple retention period that applies to every H3 API input and output in every account arrangement.
- Separate-license terms: the public community agreement says excluded-territory users should contact MiniMax, but it does not publish the terms MiniMax will offer every applicant.
Do not turn those unknowns into reassuring assumptions. “No documented upload in the workflow I inspected” is stronger than guessing that every node is offline. “The policy does not state a fixed H3 retention period” is more accurate than claiming zero retention.
MiniMax H3 Local Versus Hosted: The Short Version
| Workflow | Where generation happens | What may leave your machine | Output boundary |
|---|---|---|---|
| Local H3-Base in ComfyUI | Your computer | Initial model downloads; anything sent by optional nodes, managers, telemetry, or other tools | Native H3-Base output, normally 768px short edge |
| Local H3-Base plus MiniMax Context-IR API | Base generation is local; instruction processing is hosted | Prompt and the context needed by the hosted preprocessing request | Better-structured context for local base generation |
| Official full H3 2K workflow | Mix of local and hosted stages | Context-IR request, generated video or accessible video URL, original context, and 2K regeneration request | Up to 2K through hosted Regenerate-2K |
| OpenVeil video generation | Hosted | Active prompt and relevant input are processed by OpenVeil and necessary providers | Managed video workflow without a local GPU stack |
“Local” should describe a verified workflow, not merely a model repository.
The US License Catch Most Local-AI Guides Miss
The H3 repository is public and ungated, but public access is not the same as permission for every location and use.
The published license says:
- its applicable territory is worldwide except the EU, UK, South Korea, and United States;
- use outside that applicable territory is unauthorized under the agreement;
- commercial products or services must prominently display “MiniMax H3” in the interface;
- commercial products and services above $20 million in annual revenue need prior written authorization; and
- H3 outputs are not automatically treated as model derivatives, although the agreement still imposes use restrictions on outputs and results.
MiniMax also requires disclosure of machine-generated public content in relevant circumstances and prohibits a range of impersonation, high-risk decision, military, illegal, and rights-violating uses.
This article is not legal advice. The practical step is simple: read the current license for the exact weights you plan to use, identify where the user and deployment are located, and obtain written authorization where required before building a workflow around the model. Licenses can change; a screenshot or tutorial from release week is not a durable substitute for the current agreement.
How Much Do You Actually Need To Download?
The full Hugging Face repositories look enormous because they contain multiple model variants and alternative precisions. You do not normally download every file.
For one lower-precision FL2VA path in ComfyUI, the official Comfy-Org H3 repository currently lists approximately:
| Component | Example file | Approximate size |
|---|---|---|
| Pruned INT8 FL2VA diffusion model | minimax_h3_fl2va_pruned_int8_convrot.safetensors |
21.0 GB |
| NVFP4/AWQ text encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
15.7 GB |
| Video VAE | minimax_h3_video_vae_fp16.safetensors |
5.2 GB |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors |
0.6 GB |
| Approximate total | One FL2VA combination | 42.5 GB |
Those are decimal file sizes observed in the repository on August 8, not a promise about installed disk use. Add free space for download caching, alternate weights, the Ref2VA model if wanted, inputs, previews, temporary files, and generated videos.
The original MiniMax repository is much larger because it contains both FL2VA and Ref2VA architectures, diffusers and source layouts, full-precision components, and multiple shards. “The repository is nearly 500 GB” does not mean every user needs a 500 GB download. Conversely, “it runs on 6 GB VRAM” does not mean the whole model occupies only 6 GB on disk or in system memory.
Can MiniMax H3 Run On A Consumer GPU?
It can run on some consumer systems, but “runs” can describe very different experiences.
MiniMax's original example uses a multi-GPU inference setup. ComfyUI makes consumer use more practical with pruned and quantized weights plus dynamic model offloading. Early community reports show both the possibility and the variability:
- One user reported a four-second 320p result in 345 seconds on a laptop RTX 3060 with 6 GB VRAM and 32 GB system RAM.
- Another reported five seconds at 960×540 in 182 seconds on a laptop RTX 4090 with 16 GB VRAM.
- A 48 GB RTX 4090 environment was reported to need roughly 30 to 34 minutes for a ten-second, approximately one-megapixel result without acceleration, with shorter times after attention and prediction changes.
These are self-reported tests, not controlled benchmarks or vendor guarantees. They use different durations, resolutions, model files, workflows, and optimizations. A successful low-VRAM experiment may rely heavily on system RAM and offloading, and it may not reproduce on another 6 GB card.
Use community posts to choose a starting configuration, not to promise a client delivery time. Test the exact workflow and keep a copy of the model filenames, workflow JSON, ComfyUI version, custom-node commits, sampler settings, resolution, duration, peak VRAM, peak RAM, and wall-clock time.
Does Local H3 Keep Prompts, Faces, Voices, And Files Private?
It can keep generation inputs on the machine, but only if the entire execution path stays local.
A clean local H3-Base workflow can read a prompt and reference files from local storage, run downloaded weights on the local GPU and CPU, and write the result locally. In that arrangement, H3-Base inference does not inherently require sending the prompt or media to MiniMax.
That still leaves several places data can escape:
1. Model downloads
Hugging Face, a model manager, or another mirror can see normal download metadata such as the requested files, IP address, and request timing. The model download is not the same as uploading your prompt, but it is still network activity.
2. Custom nodes and extensions
ComfyUI workflows can include arbitrary Python nodes. A node can call a remote API, download another model, check for updates, report telemetry, or upload media. A workflow screenshot does not prove what the code does.
3. Hosted Context-IR
If you use MiniMax's Context-IR API to interpret references, the relevant prompt and multimodal context must be processed by that hosted service. MiniMax says the official hosted path also applies automated moderation to inputs and enhanced prompts.
4. Hosted 2K regeneration
MiniMax's official workflow for Regenerate-2K sends the 768p result and original context to a hosted service. Its instructions contemplate uploading a video or supplying a publicly accessible video URL. Either choice changes the privacy boundary.
5. Remote access to ComfyUI
Running ComfyUI locally does not make an exposed ComfyUI server safe. A broadly bound port, weak reverse proxy, public tunnel, shared output folder, or misconfigured cloud VM can expose prompts, reference media, workflow metadata, or generated files.
6. Local retention
Keeping data off a provider does not delete it. Input folders, browser uploads, ComfyUI history, previews, outputs, temporary files, caches, backups, Windows recent-file lists, and cloud-synced directories can retain sensitive material.
Privacy here is a system property. The weights are only one component.
A Five-Minute Locality Test
Before putting a real face, voice, private document, or unreleased product into H3, run a harmless test.
- Record the workflow. Save the JSON and list every custom node. Prefer built-in nodes when possible.
- Pre-download dependencies. Obtain the chosen model files and verify their locations and hashes where available.
- Disconnect the network. Disable Wi-Fi and unplug Ethernet after downloads finish.
- Restart the app. A workflow that succeeds only because an earlier online process cached a hosted response is not a clean offline test.
- Generate from synthetic inputs. Use a fake face, invented voice, and non-sensitive prompt.
- Inspect the result and logs. Confirm that generation completes, no node reports a failed API call, and no remote asset is silently substituted.
- Reconnect and inspect network behavior. Use an outbound firewall or network monitor to identify update checks, model-manager requests, telemetry, or API calls.
Passing that test does not prove the software has no vulnerability. It does give you evidence that the tested generation path can function without a live remote dependency.
For a broader checklist, see What To Check Before Trusting Any AI Privacy Claim.
Local H3 Versus A Hosted Privacy-Focused Workspace
Local H3 and OpenVeil solve different operational problems.
Choose a verified local H3-Base setup when:
- media must stay on hardware you control;
- you have confirmed that your location and use are licensed;
- you can manage roughly tens of gigabytes of model files plus working storage;
- you are comfortable auditing workflows and custom nodes; and
- local speed and maintenance fit the job.
Consider OpenVeil when:
- you want managed video or image-to-video tools without maintaining a local video-model stack;
- a hosted privacy-focused workspace fits the sensitivity of the task;
- you also want chat, private search, files, voice, and image tools in one place; and
- you prefer browser-local chat history rather than a normal server-side chat-history record.
OpenVeil is hosted. It is not fully offline, anonymous, or a promise of zero logs. Active prompts and relevant files are processed by OpenVeil and necessary providers to fulfill requests. OpenVeil's documented boundary is that it does not train foundation models on your prompts, uploads, images, audio, local context, or outputs; that is not the same as saying providers never process an active request.
If your requirement is “this face and voice may never leave this PC,” use a locally verified workflow and block the network. If your requirement is “I want a managed privacy-focused workspace and accept bounded provider processing,” try OpenVeil.
For the wider tradeoff, read Private AI Chat Versus Local AI and Does Using Ollama Mean Your AI Is Always Local?.
The Practical Verdict
MiniMax H3-Base is a genuine local video model, not merely a hosted API wearing an open-weight label. ComfyUI can run its released base weights locally, including workflows that generate synchronized audio. That is valuable for creators who can support the storage, memory, compute, maintenance, and audit work.
But three caveats belong in the headline, not the footnotes:
- The official complete 2K system includes hosted components.
- Consumer-GPU feasibility does not guarantee comfortable speed or quality on your hardware.
- The published community license excludes the US, EU, UK, and South Korea without separate authorization.
The privacy-safe conclusion is conditional: H3-Base can keep generation local when you use only downloaded components in an audited offline workflow, have authority under the applicable license, and secure the surrounding machine. “Open weights” alone proves none of those conditions.
Frequently Asked Questions
Is MiniMax H3 open source?
MiniMax calls H3 an open-source release, but the weights use a custom community license rather than a standard permissive open-source license. The complete system is also not fully released: Context-IR and Regenerate-2K remain hosted. “Open-weight H3-Base under a custom license” is the more precise description.
Can MiniMax H3 run fully offline?
H3-Base can run offline after the required weights, software, and workflow assets are downloaded. The official hosted Context-IR and Regenerate-2K services cannot. Optional custom nodes, model managers, remote upscalers, or telemetry can also introduce network dependencies.
Can MiniMax H3 run on 6 GB VRAM?
At least one community user reports a successful low-resolution run on a laptop RTX 3060 with 6 GB VRAM and 32 GB system RAM. Treat that as evidence of possibility, not a supported minimum. Offloading, quantization, resolution, duration, system RAM, and workflow choice can change whether it works and how long it takes.
Does local MiniMax H3 send prompts to MiniMax?
Not inherently when you run only downloaded H3-Base components in a local workflow. Prompts or media do reach a provider if you use the Context-IR API, Regenerate-2K API, or another remote node. Network monitoring and an offline test are the best practical checks.
Can I use MiniMax H3 weights in the United States?
The public community license defines the United States as an excluded territory and says use outside its applicable territory is unauthorized under that agreement. MiniMax directs excluded-territory users to request separate authorization. Review the current agreement and obtain appropriate advice for your use.
Does H3 generate audio as well as video?
Yes. The official H3-Base specification says it generates native 32 kHz stereo audio synchronized with 24 FPS video. The exact dialogue quality and language stability can vary; MiniMax lists stable support for 11 languages.
Is OpenVeil a local MiniMax H3 host?
No. OpenVeil's product documentation does not identify MiniMax H3 as a model or provider, and OpenVeil does not promise local video processing. OpenVeil is a hosted privacy-focused workspace with video and image-to-video tools, browser-local chat history, and necessary active-request processing by OpenVeil and providers.