MiniMax launched H3 on July 31, 2026, and released its open weights on August 3.
MiniMaxH3.org provides access to the full cloud-based H3 workflow, supporting text, image, video, and audio references with 4-15 second 2K video generation.
The important distinction is simple: open weights do not mean the complete H3 production pipeline can run on your own hardware.
Before choosing a workflow, most users want answers to four practical questions:
What exactly did MiniMax release?
Can MiniMax H3 run locally?
Can local H3 generate 2K video?
Is local deployment actually cheaper than using H3 in the cloud?
At a high level, local H3 favors deeper control, while cloud H3 favors faster, production-ready video creation.
Is MiniMax H3 Fully Open Source?
The core weights are open; the full production stack is not.
MiniMax describes H3 as open source and has released the model checkpoints under the MiniMax H3 Community License Agreement. The release includes the core H3 generation components required for local inference.
However, the complete production workflow still has three distinct stages:
H3 Component | What It Does | Current Availability |
H3-Context-IR | Organizes complex text, image, video, and audio references | Hosted / Cloud support |
H3-Base | Generates synchronized video and audio at 768P | Open weights |
H3-Regenerate-2K | Regenerates video at 2K using the original context | Hosted for now / Cloud support |
This structure is the clearest way to understand the release:
the central generation engine is available locally, while important input-processing and high-resolution stages remain hosted.
MiniMax documentation also states that H3-Context-IR is not included in the open release, and H3-Regenerate-2K has not yet been released as an open component.
In practice, the local H3 release is not the complete H3 experience, so quality, efficiency, and consistency may fall short of the full cloud workflow.
What Can You Run Locally Actually?
Local H3 includes the core generation engine.
MiniMax currently provides two task-specific H3-Base checkpoints.
MiniMax-H3 Base FL2VA supports:
Text-to-video with audio
First-frame-to-video
Last-frame-to-video
First-and-last-frame control
MiniMax-H3 Base Ref2VA focuses on:
Reference-to-video generation
Text combined with image, video, and/or audio references
Both checkpoints generate video and native audio together rather than requiring a separate audio-generation step. The full H3 model supports 24 FPS output and 32 kHz stereo sound.
These local checkpoints are useful for developers building:
Custom AI video workflows
Reference-driven generation systems
Character and motion tools
Internal creative pipelines
Research and experimental applications
Local capability is real, but for ordinary creators, supported locally and easy to operate locally are still very different things.
What Hardware Does Local Deployment Require?
The weights are accessible; the compute is not lightweight.
MiniMax H3 is a substantial video model rather than a lightweight desktop tool.
MiniMax's official SGLang deployment example uses four GPUs for both FL2VA and Ref2VA. That should not be treated as a universal four-GPU minimum, because requirements can change with frameworks, quantization, and future optimizations, but it shows the scale of the native workload.
The official repository is also large, although users can download only the task families and components required by their chosen inference framework rather than every available file.
Running H3 locally also means managing:
GPU memory and compute
Model downloads and storage
CUDA and framework dependencies
Inference configuration
Updates and optimization
Failed jobs and debugging
Server availability
Production maintenance
MiniMax also notes that the first open release uses full-attention inference, while its sparse-attention implementation is planned for a later release.
The weights may be available, but the infrastructure remains your responsibility.
Can Local MiniMax H3 Generate 2K Video?
Local H3 stops at 768P; 2K still depends on hosted components.
The open-weight H3-Base currently generates 768P video. MiniMax's official 2K pipeline then adds H3-Regenerate-2K, which uses the 768P result together with the original context to regenerate the video at 2K. That regeneration module is not open yet.
The practical difference is straightforward:
Local H3: Best suited to 768P experimentation, research, and custom workflows.
Cloud H3: Provides a much easier path to 2K AI video without managing the regeneration pipeline, GPUs, or local infrastructure.
MiniMax H3's 2K process is also more than conventional upscaling. It reuses the original multimodal context during regeneration, helping preserve finer visual details that simple enlargement could lose.
For local control, H3-Base is the starting point. For fast, production-ready 2K output, cloud generation is currently the simpler route.
Try MiniMax H3 2K Video Generator
MiniMax H3 Local vs Cloud: Which Costs Less?
Local shifts cost to infrastructure; cloud turns it into a per-video expense.
Local H3 can become economical at scale, but it is not free in practice. You still need to account for GPUs, storage, setup, maintenance, electricity, and engineering time.
Teams that already operate suitable infrastructure and generate large volumes of video can spread those fixed costs across many jobs.
Cloud H3 uses a different cost model: you pay per generation or through credit packs, while most of the technical overhead is handled for you.
Cost Factor | Local H3 | |
Model access | Open weights | Included |
GPU hardware | Required | Not required |
Setup & maintenance | Self-managed | Managed |
2K production | More complex | Easier |
Cost model | Infrastructure-based | Per generation / credits |
Best for | High-volume technical teams | Creators, marketers, agencies |
Current MiniMax H3 cloud plans start at $9.90. New users receive 90 free credits after signing in, while a 15-second 2K generation costs about $3.78–$5.62 depending on the credit pack.
For most individual creators and small production teams, cloud H3 is therefore the simpler starting point. Local deployment becomes more attractive when generation volume is high, infrastructure already exists, or deeper customization is required.
Local H3 can reduce marginal cost at scale; cloud H3 reduces the cost and complexity of getting started.
Pricing note: Actual costs can change over time and may vary by platform, credit package, resolution, duration, reference assets, failed generations, promotions, infrastructure costs, and future model updates. Always check current pricing before budgeting.
Can You Run MiniMax H3 in ComfyUI?
ComfyUI improves control, not hardware efficiency.
ComfyUI provides local MiniMax H3 workflows for:
Text-to-Video
Image-to-Video
Reference-to-Video
Native stereo audio generation
However, ComfyUI is not the same as a lightweight web generator. You still need compatible model files, GPU resources, storage, and a stable local environment.
For advanced users, that flexibility can be valuable. For creators who mainly want to turn prompts and references into finished videos, a managed cloud workflow remains much simpler.
What Should U.S. Users Know About the License?
Hosted access is simpler; local deployment needs an extra license check.
The current MiniMax H3 Community License lists the United States as an Excluded Territory, together with the EU, UK, and South Korea. That means the standard open-weight license does not automatically authorize local deployment in the U.S.; organizations in those regions can apply to MiniMax for separate permission.
The key distinction is:
Open-weight / local H3: U.S. deployment requires reviewing the current license and may require separate authorization.
Hosted H3 / API: MiniMax states that its API remains globally available because the hosted service includes its own safety and compliance controls.
MiniMax says these restrictions reflect the evolving regulatory environment around generative video, including copyright, likeness generation, content safety, and responsible deployment. The company also describes the current limitation as temporary rather than permanent.
For most U.S. creators, marketers, and production teams, cloud-based H3 is therefore the simpler path because there is no local deployment setup or open-weight licensing process to manage directly.
This is a summary of the current MiniMax license, not legal advice.
Can MiniMax H3 Be Used for Commercial Work?
Commercial use depends on territory, scale, and license terms.
Within applicable territories, the Community License grants rights to use, modify, reproduce, distribute, and create model derivatives subject to its conditions and Acceptable Use Policy. MiniMax also states that it claims no rights over generated outputs.
The license includes additional commercial conditions. For example, separate prior written authorization is required if commercial products or services using MiniMax H3 generate more than $20 million in yearly revenue. Commercial products using the H3 works must also prominently display “MiniMax H3” in their interface.
For U.S. organizations, the territorial restriction discussed above should be reviewed first.
This is why teams should distinguish between using H3 through a hosted cloud service and deploying the H3 model weights themselves.
What Are the Main Limitations of Running Locally?
Local H3 gives you more control, but the current open release still falls short of a full production workflow.
A few limitations become important when you move beyond testing and start using H3 for real projects:
2K is not fully local. H3-Base is centered on 768P generation, while cloud workflows make 2K output much easier by handling the hosted regeneration stage for you.
Local inference is still under-optimized. The current open release does not yet include sparse-attention support, so local performance has not reached its full efficiency potential.
Generation modes are still separate. Image-to-video and multimodal reference-to-video follow different workflows, adding complexity to advanced local pipelines.
You manage compliance yourself. Watermarking, AI-content labels, and publishing requirements may need to be configured based on your platform and market.
For developers, that extra control may justify the added complexity. For creators, marketers, and agencies, cloud H3 offers a simpler path with less setup, maintenance, and 2K workflow management.
Local H3 is built for control; cloud H3 is built for production efficiency.
MiniMax H3 Local or Cloud: Which Workflow Fits You?
Choose local for control. Choose cloud for production.
Neither approach is universally better. The right choice depends on your infrastructure, technical goals, and production priorities.
Local H3 | ||
Setup | Technical | Minimal |
Local GPU | Required | Not required |
Infrastructure | Self-managed | Managed |
Customization | High | Lower |
Easy 2K production | More complex | Yes |
Maintenance | Your team | Provider |
Cost model | Hardware + operations | Credits / generation |
Best for | Developers, research, custom systems | Creators, marketers, agencies |
Choose MiniMax H3 local if:
You already have capable GPU infrastructure.
You need deeper model or workflow customization.
You are developing research, internal tools, or custom AI pipelines.
Keeping inference on your own infrastructure is important.
Your use complies with the current MiniMax H3 license.
Choose MiniMax H3 cloud if:
Your priority is simply to create videos faster.
You want to avoid GPU setup, model downloads, and maintenance.
You need easier access to 2K output.
You create ads, music visuals, product videos, game content, or social media assets.
You prefer predictable generation costs over managing infrastructure.
For individual creators and smaller production teams, paying per generation is often more practical than building and maintaining a dedicated local H3 environment.
Why Does the Open Release Matter Even for Cloud Users?
Open weights grow the H3 ecosystem, even if you never run the model locally.
The open release is valuable because it gives developers more room to improve and extend H3 beyond the default workflow.
That can lead to:
Faster inference methods
Better ComfyUI integrations
Custom creative and production tools
Research and testing workflows
Purpose-built interfaces
New multimodal and reference-based video pipelines
Cloud users can benefit from these improvements without managing GPUs, model files, or local deployment themselves.
Developers can extend H3; creators can use those advances through simpler cloud workflows.
Choose local H3 when you need control. Choose cloud H3 when you need faster production.
Local deployment is better suited to developers, researchers, and teams building custom pipelines or managing their own infrastructure.
For creators, marketers, designers, musicians, and production teams, MiniMax H3 cloud is usually the easier starting point
no model downloads
no GPU setup
a more direct path from creative references to finished 2K video
Want to create instead of configure?
MiniMax H3 Local vs Cloud FAQs
Is MiniMax H3 free to download?
Yes. H3-Base weights are publicly available, but local use still requires GPU resources and compliance with the current license.
Is MiniMax H3 fully open source?
Not entirely. H3-Base is open, while H3-Context-IR and H3-Regenerate-2K are not currently part of the open release.
Can U.S. users run MiniMax H3 locally?
Local deployment has extra license restrictions in the U.S. The current Community License lists the United States as an Excluded Territory, so separate authorization may be required.
Is cloud H3 cheaper than local deployment?
It depends on usage. Local can suit teams with existing GPUs and high volume, while cloud is often simpler for creators and small teams.
MiniMaxH3.org credit packs start at $9.90, and new users receive 90 free credits after signing in.