H3 Base weights
Verified released
Complete H3-Omni-Transformer weights are publicly downloadable.
Open-source status · Verified August 21, 2026
Is MiniMax H3 open source?
H3 Base weights are public. The complete production system is not fully local.
Short answer: MiniMax has released downloadable H3 Base weights, inference code, and two task-specific checkpoints. H3-Context-IR and H3-Regenerate-2K remain hosted, while native sparse-attention code is still pending. The files use the custom MiniMax H3 Community License—not a standard permissive open-source license.
Last verified August 21, 2026 against MiniMax’s official GitHub, Hugging Face, launch post, license, and API documentation. Check the linked originals before downloading or deploying.
Release map
What is available now
The later GitHub and Hugging Face artifacts—not the launch phrase “open model” alone—establish what can actually be downloaded.
FL2VA checkpoint
Verified released
Text-to-audio-video with optional first and last frames at 768p.
Ref2VA checkpoint
Verified released
Text plus image or video references, with optional audio, at 768p.
Inference code
Verified released
Processors, transformer, VAEs, schedulers, examples, and configs.
H3-Context-IR
Not released for local use
Not included for local use; MiniMax provides it through an API.
H3-Regenerate-2K
Not released for local use
Not locally released; the official 2K path still uses its API.
Native sparse attention
Not included in initial release
Not included in the initial release; local inference uses full attention.
Fully offline complete 2K workflow
Not verified
The official full workflow still depends on hosted Context-IR and Regenerate-2K services.
“Released” means an official public artifact was downloadable when checked. It does not mean every H3 service is local, the license is OSI-approved, or that we independently reproduced MiniMax’s results.
License checkpoint
Public weights, with important limits.
MiniMax distributes H3 under the MiniMax H3 Community License Agreement, dated August 2, 2026. It is a custom license—not Apache 2.0, MIT, or another standard permissive software license.
The agreement defines an Applicable Territory that excludes the European Union, United Kingdom, Republic of Korea, and United States. It also includes use, distribution, attribution, and commercial conditions, including separate written authorization above its stated annual-revenue threshold.
Read the current agreement before use or redistribution. This is an orientation summary, not legal advice or permission.
Deployment paths
Local 768p is not fully local 2K
MiniMax documents local H3 Base and a separate hybrid path for the complete 2K workflow.
01
Hosted workflow
Use the generator on this site or MiniMax’s API. No local model setup is required.
02
Local 768p H3 Base
Follow the official SGLang, vLLM, diffusers, or ComfyUI instructions after checking the license.
03
Hybrid 2K workflow
Run H3 Base locally while using the hosted Context-IR and Regenerate-2K APIs.
The official repository lists SGLang, vLLM, diffusers, and ComfyUI. Its SGLang example uses four GPUs, but this is not a universal minimum. Requirements change with checkpoint, framework, precision, offloading, resolution, and workload.
For prompt structure before choosing a runtime, use our MiniMax H3 prompt guide. We do not claim that any named consumer GPU is sufficient.
Evidence
What we verified—and what we did not
Verified from first-party artifacts
- An official public GitHub repository and non-gated Hugging Face model exist.
- FL2VA and Ref2VA checkpoint families are downloadable.
- The release documents local 768p and a hybrid path to 2K.
- Context-IR, Regenerate-2K, and native sparse-attention code are not in the initial local release.
- The model uses the custom MiniMax H3 Community License.
Not independently verified by MiniMaxH3.xyz
- Output quality, speed, memory use, or hosted-service parity.
- A universal minimum GPU or consumer hardware configuration.
- Successful installation through every linked framework.
- A fully offline 2K workflow equivalent to the hosted system.
- Legal suitability for a reader’s country or intended use.
For product capability context, read our evidence-led MiniMax H3 review.
Primary sources
Check the originals
Every source below was last checked August 21, 2026. Repository contents and terms can change.
MiniMax H3 GitHub repositoryhttps://github.com/MiniMax-AI/MiniMax-H3Weights, code, frameworks, and local/hybrid workflow status2026-08-21MiniMax H3 model on Hugging Facehttps://huggingface.co/MiniMaxAI/MiniMax-H3Public model files and checkpoint artifacts2026-08-21MiniMax H3 Community Licensehttps://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSECurrent license terms and restrictions2026-08-21MiniMax video generation docshttps://platform.minimax.io/docs/guides/video-generationHosted API modes and the 2K service path2026-08-21
FAQ
Questions about the release
Is MiniMax H3 open source?
MiniMax has released H3 Base weights and inference materials under its custom Community License. Parts of the production system remain hosted, so “open-weight release under a restricted community license” is more precise than implying the complete system uses a standard permissive open-source license.
Can I download the MiniMax H3 model weights?
Yes. The official Hugging Face repository exposes FL2VA and Ref2VA checkpoint families, and the GitHub repository provides framework guidance. Read the current Community License before downloading, using, or redistributing them.
Can MiniMax H3 run completely offline?
The released H3 Base checkpoints support the documented local 768p path. The complete official 2K workflow is not fully offline because it combines local H3 Base with hosted H3-Context-IR and H3-Regenerate-2K APIs.
Does MiniMax H3 work with ComfyUI?
MiniMax’s official repository links to ComfyUI guidance and T2V/R2V templates. MiniMaxH3.xyz has not independently installed or benchmarked that workflow, so use the current official instructions rather than an unverified copied workflow.
What GPU do I need for local MiniMax H3?
MiniMax’s SGLang example uses four GPUs, but it is not presented as a universal minimum. Requirements vary by checkpoint, framework, precision, offloading, resolution, and workload. We do not claim that a particular consumer GPU is sufficient.
Can I use MiniMax H3 commercially?
The Community License describes commercial use but includes territorial limits, use restrictions, attribution requirements, and an authorization condition above its stated revenue threshold. Read the current license and obtain qualified legal advice; this page does not grant permission.
Choose your route
Run online, or inspect the release.
Generate through the hosted workflow without downloading local weights—or start from MiniMax’s official repository.