Model weights
Available
MiniMax publishes two BF16 base task checkpoints on Hugging Face.
Open-weight status checked · 5 August 2026
MiniMax H3 Open Source? Open Weights, Licence, and Local Use
MiniMax H3 open weights are real, downloadable, and supported by local inference tools. But the complete H3 product is not fully released, and the Community License has major territory and commercial conditions. This independent guide shows exactly what MiniMax released, what remains hosted, and when a local model deployment makes sense.
The standard licence does not authorise use in the EU, UK, Republic of Korea, or United States. This page is informational and is not legal advice.
MiniMax H3 release map
Weights ≠ complete hosted system
H3-Base
FL2VA + Ref2VA checkpoints · local 768p · stereo audio
Context-IR
Hosted
Regenerate-2K
Not released
Custom Community License
The precise answer
Is MiniMax H3 open source or open weights?
MiniMax H3 is best described as an open-weight video model with a partially released system. MiniMax publishes the H3-Base FL2VA and Ref2VA checkpoints, model configuration, inference guidance, and components needed for local base generation. Developers can download the weights, inspect the released structure, run H3 locally, and build derivatives when the Community License authorises their location and use.
“MiniMax H3 open source” is a popular search phrase, but it can imply more than the release provides. The licence is not a standard OSI-approved open-source licence: it limits geography, restricts certain uses of H3 outputs, and adds commercial and downstream-service duties. The complete MiniMax pipeline also includes hosted or unreleased modules. Open weights answer “can I obtain and run the parameters?”; full open source would additionally require broadly reusable code and rights across the complete system without these field and territory limits.
The distinction is practical, not semantic nit-picking. A researcher in an authorised territory may use the released checkpoints for local experiments. A company in an excluded territory cannot assume that downloading the same MiniMax files grants deployment rights. A creator using the global API is using a hosted MiniMax service governed by separate terms, not exercising the same rights as an open-weight licensee.
Local output
768p base
The local base model produces video and native stereo audio at 24 fps.
Complete system
Not fully released
Context-IR, 2K regeneration, and sparse attention are not all released.
Licence
Custom terms
Territory, commercial, attribution, and acceptable-use conditions apply.
Available now
What the MiniMax H3 public release includes
The initial MiniMax release focuses on H3-Base. Each released checkpoint is a self-contained task family with the processor, tokenizer, Qwen-based text encoder, H3 transformer, visual VAE, and audio VAE needed for audiovisual inference.
1
H3-Base FL2VA
The first MiniMax checkpoint supports text-to-audio-video plus first-frame, last-frame, and first-and-last-frame generation. It is distributed in BF16 with the processor, tokenizer, encoder, transformer, visual VAE, and audio VAE required by the task family.
2
H3-Base Ref2VA
The second MiniMax checkpoint accepts text with reference images, videos, and audio. It is designed for identity, style, motion, camera, and voice guidance while generating synchronized picture and stereo sound.
3
Inference integrations
The MiniMax model card points to SGLang, vLLM, Diffusers, and ComfyUI. ComfyUI also publishes native templates and repackaged files for text-to-video, image-to-video, and reference-to-video workflows.
4
Development materials
MiniMax publishes model configuration, prompt-writing guidance, sample scripts, validation examples, and full base-model weights. Those materials make research and workflow development possible within the licence scope.
Duration
4–15 seconds
Frame rate
24 fps
Base canvas
768px short edge
Audio
32 kHz stereo
Not in the package
What remains closed or incomplete
MiniMax describes H3 as a three-part system: Context-IR, H3-Base, and Regenerate-2K. Only treating the base weights as “the whole model” hides important quality and deployment differences. A local pipeline must replace the unavailable preprocessing logic, accept base-resolution output, or connect to hosted MiniMax endpoints.
H3-Context-IR
This hosted preprocessing and orchestration layer interprets free-form multimodal inputs and turns them into the structured context consumed by the base model. MiniMax says its internal multi-stage workflow is not included in the public release.
H3-Regenerate-2K
The MiniMax 2K path feeds the base result and original context back through the model. That regeneration module is not yet available, so a fully local H3-Base workflow should not be advertised as the same end-to-end 2K system.
Native sparse attention
The first MiniMax release uses full-attention inference. MiniMax says the sparse-attention implementation used during the final training stage will be released separately in a future update.
Architecture
Inside the open MiniMax H3 Base model
MiniMax H3 uses a single audiovisual generation path rather than creating silent video and attaching sound later. Understanding the released architecture also explains why MiniMax BF16 deployment is substantial and why a small “minimum VRAM” claim would be misleading.
Qwen3-VL-32B encoder
H3 uses the full pretrained encoder weights and passes hidden states from layer 50 into the Omni Transformer. MiniMax adds required special tokens and configuration.
33B dense transformer
The H3 Omni Transformer jointly predicts visual and audio latents. About 13B parameters sit in AdaLN-related branches that can be precomputed for inference-only deployment.
Visual VAE
The MiniMax video autoencoder compresses space and time before patchification, reducing the visual sequence processed by the transformer.
Stereo Audio VAE
MiniMax keeps separate left and right channels on the same encoder and decoder, then recombines them into synchronized 32 kHz stereo output.
Full attention first
The initial release uses full-attention inference. The native sparse implementation described by MiniMax is not yet included.
Two task checkpoints
FL2VA covers text and keyframe modes; Ref2VA covers multimodal references. They are distinct open weight sets rather than one interchangeable file.
Licence checkpoint
MiniMax H3 open weights come with major conditions
The MiniMax H3 Community License took effect on 2 August 2026. It grants limited rights only inside the “Applicable Territory,” defined as the world excluding the EU, UK, Republic of Korea, and United States. The restrictions apply not only to the H3 weights but also to defined H3 works, model derivatives, and outputs. A public download link does not override those terms.
Territory
The standard open licence excludes four major regions. MiniMax says organisations there may contact it for a separate licence with compliance controls and guardrails.
Commercial use
Commercial products must prominently display MiniMax H3. Above USD 20 million in yearly product or service revenue, prior written authorisation is required.
Hosted safeguards
A third-party MiniMax service must bind users to protective terms, implement safeguards, provide a reporting route, investigate reports, and address repeat violations.
Model improvement
The licence restricts using H3 works, outputs, or results to improve another AI model, except MiniMax H3 or qualifying H3 model derivatives.
Redistribution
Redistribution inside the applicable territory requires the licence, modification notices, and a NOTICE file. Downstream users must receive applicable restrictions.
Outputs
MiniMax says it claims no rights over generated outputs, but users remain responsible for those outputs and must satisfy territory, disclosure, safety, and rights obligations.
Do not rely on a summary for deployment approval
Read the current official agreement and its Acceptable Use Policy, confirm the physical and legal locations of developers, servers, users, outputs, and downstream services, then obtain qualified legal advice. MiniMax can update policies, and project-specific facts may change the analysis.
Open video comparison
MiniMax H3 vs other open video models
“Open” covers different combinations of weights, code, training support, licences, and hosted services. This comparison focuses on release structure and deployment fit rather than unverified quality rankings. Test every model on the same prompts, references, duration, resolution, and hardware.
| Model | Open access | Audio approach | Licence | Best evaluation fit |
|---|---|---|---|---|
| MiniMax H3 | Open weights plus local inference materials | Native stereo audio generated with video | MiniMax H3 Community License; excluded territories and additional terms | Multimodal reference control and audiovisual generation where the licence permits deployment |
| LTX-2 | Public model, inference, pipeline, and trainer repositories | Joint audio-video diffusion architecture | LTX-2 Community License; review commercial-use conditions | Local audiovisual workflows, training experiments, and an established LTX tool ecosystem |
| Wan 2.2 | Public code, weights, and multiple task variants | Audio-driven S2V exists; capabilities vary by checkpoint | Apache 2.0 for the official repository models | Broad video-generation experimentation with a permissive repository licence |
| HunyuanVideo 1.5 | Public weights, inference, training, and LoRA materials | Video-focused rather than native joint stereo output | Tencent Hunyuan Community License with territorial restrictions | A lighter video model when its licence territory and workflow match the project |
Licences and repositories can change. “Open access” does not mean unrestricted commercial use, global availability, identical output rights, or an OSI-approved licence.
Local deployment
How to evaluate MiniMax H3 open weights
Start with legal eligibility, then choose the smallest H3 path that answers your product question. The original BF16 checkpoints are large, while quantized and repackaged variants trade memory, speed, and sometimes fidelity. Record every file, commit, precision, node version, resolution, duration, attention implementation, and offload setting.
- 1
Verify licence scope
Confirm territory, commercial status, planned distribution, output location, downstream users, and acceptable-use controls before downloading or running H3.
- 2
Choose one open task family
Use FL2VA for text and keyframes or Ref2VA for multimodal reference generation. Do not load both weight sets unless the test requires both.
- 3
Select an official integration
Follow MiniMax guidance for SGLang, vLLM, Diffusers, or the native ComfyUI workflow. Pin versions and verify file checksums.
- 4
Run a base-resolution control
Start with a short 768p MiniMax clip and simple synchronized sound. Confirm identity, motion, dialogue, effects, music, and final MP4 metadata.
- 5
Measure the complete system
Track download size, peak VRAM, system RAM, load time, render time, failures, output consistency, and any quality change from quantisation or acceleration.
Decision guide
When MiniMax H3 open weights are a good fit
Consider H3 when
- Your users, infrastructure, and outputs remain inside an authorised territory.
- You need open local audiovisual weights with native stereo sound.
- You value image, video, audio, identity, motion, camera, and voice references.
- You can operate a large model and accept that the full hosted system is not open.
Choose another path when
- Your project operates in an excluded territory without separate authorisation.
- You require an OSI-approved, globally permissive licence.
- You need the exact closed Context-IR or Regenerate-2K workflow locally.
- Your hardware, compliance controls, or review process cannot support the model responsibly.
FAQ
MiniMax H3 open-source questions
Is MiniMax H3 open source?
MiniMax H3 has open weights, model configuration, inference integrations, and supporting documentation, but calling the complete H3 system fully open source is imprecise. The custom licence restricts territory and use, while Context-IR, Regenerate-2K, and native sparse attention are not all part of the initial release.
Are MiniMax H3 open weights available now?
Yes. The official Hugging Face repository includes H3-Base FL2VA and H3-Base Ref2VA checkpoints. The model card documents original and Diffusers-style layouts, while Comfy Org provides files and native ComfyUI templates for local workflows.
Can MiniMax H3 open weights be used in the US or EU?
Not under the standard August 2, 2026 Community License. The licence defines the European Union, United Kingdom, Republic of Korea, and United States as excluded territories. MiniMax invites interested users in those locations to contact the company about a separate licence. This is not legal advice; obtain qualified advice for your deployment.
Can MiniMax H3 be used commercially?
Commercial use is possible only within the applicable territory and subject to the full licence. Commercial products must prominently display MiniMax H3, and products or services generating more than USD 20 million in yearly revenue require prior written authorisation. Other distribution, safeguards, and acceptable-use obligations also apply.
Does local MiniMax H3 generate 2K video?
The open H3-Base workflow is documented for 768p generation. The official 2K result uses H3-Regenerate-2K, which is not yet open. A local workflow may upscale or use other processing, but it should not claim to reproduce the unreleased MiniMax regeneration module.
What hardware is required for MiniMax H3?
MiniMax does not publish one universal minimum in the model card. The MiniMax system combines a 33B dense Omni Transformer, a Qwen3-VL-32B encoder, and separate visual and audio VAEs, so the original BF16 release is substantial. Quantized ComfyUI files can change memory requirements, speed, and quality. Benchmark the exact checkpoint, resolution, duration, precision, and offload strategy on your hardware.
Can the H3 outputs train another AI model?
The standard licence says MiniMax H3 works, outputs, and results may not be used to improve another AI model except MiniMax H3 or its model derivatives. Review the exact definition of model derivatives and obtain legal advice before dataset generation, distillation, or cross-model training.
Is the MiniMax API covered by the open-weight licence?
MiniMax distinguishes hosted API access from independently deployed open weights. Its licence Q&A explains that the company can operate compliance controls for the API, while downloaded weights leave its infrastructure. Check the current API terms separately from the Community License that governs the released H3 works.
Verify the release
Read the MiniMax H3 open documents
The MiniMax model card and licence are the controlling starting points for any H3 evaluation. Review them again before deployment because open files, supported runtimes, policy explanations, and model modules may change after this page is published.
Next Vibe AI is independent and is not affiliated with or endorsed by MiniMax.
Compare MiniMax H3 with LTX