MiniMax
MiniMax H3
MiniMax H3 is an open-weight audio-video family with local 768p H3-Base checkpoints and a separate hosted path for Context IR and 2K regeneration.
Capabilities
Local deployment evidence
MiniMax’s official model repository publishes BF16 H3-Base FL2VA and Ref2VA weights together with local SGLang commands.
Hardware entry summary
No documented numeric hardware entry value
View full hardware profiles ↓Local checkpoints
MiniMax H3-Base FL2VA
Variant ID: minimax-h3-base-fl2va
- Access mode
- Local checkpoint
- Checkpoint type
- Distilled
- Precision
- BF16
- Maximum duration
- 15 seconds
- Maximum resolution
- 768p
- Frame rates
- 24 fps
- Aspect ratios
- 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
- Native audio
- Yes
- Reference image
- Yes
- Pricing
- No provider inference fee
Notes
Official CFG-distilled BF16 checkpoint for text-to-audio-video and first/last-frame-conditioned audio-video generation. Local execution has no provider inference fee, but users supply GPUs, storage, electricity, and time. Use remains subject to the MiniMax H3 Community License.
MiniMax H3-Base Ref2VA
Variant ID: minimax-h3-base-ref2va
- Access mode
- Local checkpoint
- Checkpoint type
- Distilled
- Precision
- BF16
- Maximum duration
- 15 seconds
- Maximum resolution
- 768p
- Frame rates
- 24 fps
- Aspect ratios
- 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
- Native audio
- Yes
- Reference image
- Yes
- Pricing
- No provider inference fee
Notes
Official CFG-distilled BF16 checkpoint for reference-conditioned audio-video generation from text plus image, video, or audio references. Reference video is not represented as conventional video-to-video capability without an official one-to-one mapping. Local costs and license conditions still apply.
Hardware profiles
Official four-GPU SGLang example
- Applies to
- MiniMax H3-Base FL2VA, MiniMax H3-Base Ref2VA
- Official minimum VRAM
- Unknown
- Official recommended VRAM
- Unknown
- Official minimum system RAM
- Unknown
- Official recommended system RAM
- Unknown
- Storage
- Unknown
The official SGLang examples use four GPUs with Ulysses degree 4 for both checkpoints. No per-GPU VRAM, system RAM, or storage minimum is stated, so no numeric requirement is inferred.
Official documentation (External link)Overview
- Provider
- MiniMax
- Status
- Active
- Publication status
- Published
- Record scope
- Family
- Release date
- 2026-08-02
- Last verified
- 2026-08-06
Specifications
- Family maximum duration
- 15 seconds
- Family maximum resolution
- Varies by variant
- Frame rates
- 24 fps
- Aspect ratios
- 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
- Portrait support
- Yes
- Native audio
- Yes
- Reference image
- Yes
- Character consistency
- Unknown
Hardware
- Local deployment
- Yes
- Tested minimum VRAM
- Unknown
- ComfyUI support
- Yes
Licensing
- Open weights
- Yes
- Open source
- Unknown
- Commercial use
- Conditional
- License
- MiniMax H3 Community License Agreement
Availability
- API availability
- Yes
- Web app availability
- Yes
Performance and pricing
- Speed category
- Unknown
- Estimated clip cost
- Unknown
- Pricing
- Varies by variant
Hosted API options
MiniMax H3 full 2K hybrid workflow
Variant ID: minimax-h3-full-2k-hybrid-api
- Access mode
- Hosted API
- Checkpoint type
- Not applicable
- Precision
- Not specified
- Maximum duration
- 15 seconds
- Maximum resolution
- 2K
- Frame rates
- 24 fps
- Aspect ratios
- 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
- Native audio
- Yes
- Reference image
- Yes
- Pricing
- Pricing unknown
Notes
The documented 2K workflow is hybrid, not a downloadable 2K checkpoint: local H3-Base produces 768p, while hosted H3-Context-IR preprocessing/orchestration and H3-Regenerate-2K complete the online path. API pricing is not normalized in this record.
Best use cases
- Local text- or reference-conditioned 768p audio-video generation
- Hybrid workflows that combine local H3-Base output with MiniMax’s hosted 2K path
Limitations
- The official local examples use four GPUs but do not state per-GPU VRAM or system-memory requirements.
- 2K regeneration and Context IR are hosted services, not released local weights.
- The community license has territorial, revenue, attribution, redistribution, safeguards, and model-improvement restrictions.
Hardware notes
The official local command is a four-GPU SGLang example. Because the provider does not state a per-GPU VRAM, system RAM, or storage minimum, all numeric hardware fields remain unknown.
Commercial use notes
Commercial use is conditional. The license excludes the EU, United Kingdom, Republic of Korea, and United States; organizations with more than USD 20 million in annual revenue need prior written authorization for commercial products or services. Commercial interfaces must prominently display “MiniMax H3”; redistribution must carry the license and NOTICE; safeguards and use restrictions apply; and H3 materials or outputs may not be used to improve a non-H3 AI model. Review the complete license before use.
Speed notes
No Aenica reproducible benchmark has been completed, so speed remains unknown.
Pricing notes
Local H3-Base checkpoints have no provider inference fee but still require user-funded compute, electricity, storage, and time. Hosted Context IR, API, and 2K regeneration pricing is not normalized here.
Sources
- Official documentationMiniMax H3 official model repository (External link)
- Official documentationMiniMax H3 FL2VA official weights folder (External link)
- Official documentationMiniMax H3 Ref2VA official weights folder (External link)
- Official documentationMiniMax H3 official repository (External link)
- Official documentationMiniMax H3 video generation API guide (External link)
- Official documentationComfyUI MiniMax H3 official integration guide (External link)
- Official announcementMiniMax H3 release announcement (External link)
- Official licenseMiniMax H3 Community License Agreement (External link)