What Is the MiniMax H3 Video Model?
MiniMax H3 is a general-purpose multimodal video generation model developed by MiniMax. It can understand context across text, images, video, and audio and generate audiovisual video with native stereo sound. Current H3 workflows support 4–15 second output at 768p or up to 2K. The open-weight release includes H3-Base in two task families — FL2VA (First-Last-frame-to-Video-and-Audio) and Ref2VA (Reference-to-Video-and-Audio).
- Inputs: text prompt, image references (first and/or last frame), video references, audio references.
- Output: MP4 video with optional synchronized stereo audio, 4–15 seconds, 768p or 2K.
- Cost on MiniMax3.org: 6 credits/second at 768p, 12 credits/second at 2K.
- Powered by: Hailuo AI (MiniMax's consumer video product) runs on MiniMax H3.
MiniMax H3 Model Specifications
| Attribute | Value |
|---|---|
| Model name | MiniMax H3 |
| Developer | MiniMax |
| Category | Video generation model (multimodal) |
| Input modalities | Text prompt · Image reference (first and/or last frame) · Video reference · Audio reference |
| Output | Video clips (MP4) · Optional synchronized audio |
| Resolutions | 768p, 2K |
| Duration range | 4–15 seconds |
| Audio support | Optional generated audio on supported workflows |
| Main use cases | Text-to-video, Image-to-video, First/last frame video, Reference-guided video |
| Access | Web (minimax3.org), API, ComfyUI, Hugging Face |
| Open weights | H3-Base (FL2VA & Ref2VA) on Hugging Face and GitHub |
| Last verified | 2026-08-11 |
Specifications sourced from MiniMax official documentation and Hugging Face model card. Last verified 2026-08-11.
MiniMax H3 System Architecture
The complete MiniMax H3 generation system has three major stages with different responsibilities. Understanding the architecture helps clarify what the open weights provide versus the full hosted system.
H3-Context-IR
Interprets complex multimodal instructions and relationships across text, images, video, and audio. Converts context into an intermediate representation that H3-Base can use reliably.
H3-Base
Performs core audiovisual generation and produces the base 768p result. The open-weight release includes FL2VA and Ref2VA task families. This is what you get when you download MiniMax H3.
H3-Regenerate-2K
Takes the base result together with the original context and regenerates the output at 2K. Part of the complete H3 system, not a separate user-facing model.
Open H3-Base vs the Full MiniMax H3 System
| Capability | Open H3-Base | Complete H3 System |
|---|---|---|
| Core audiovisual generation | Yes | Yes |
| 768P generation | Yes | Yes |
| FL2VA | Yes | Yes |
| Ref2VA | Yes | Yes |
| H3-Context-IR | Not included in open weights | Available as part of the full workflow |
| H3-Regenerate-2K | Not part of H3-Base alone | Used for full 2K workflow |
| Local deployment | Yes | Hybrid/hosted components required for complete workflow |
If you download MiniMax H3 locally, you are primarily deploying H3-Base. For full 2K and enhanced context processing, use the API or MiniMax3.org's hosted generator above.
MiniMax H3 Key Capabilities
Multimodal Conditioning
MiniMax H3 accepts four input modalities simultaneously: text prompts, image references (including first and last frame), video references, and audio references. Each modality can carry separate roles — a portrait image for identity, a video clip for motion, and an audio track for voice or music. This multimodal conditioning is what separates H3 from text-only video models.
Native Audio-Visual Generation
Unlike most video models that generate silent footage, MiniMax H3 can generate native, synchronized stereo audio alongside video. Audio generation works with voice, music, and ambient sound references, producing output where visuals and sound are temporally aligned from the model itself — not post-processed.
High-Resolution Output up to 2K
H3-Base generates at 768p. The complete H3 system adds an H3-Regenerate-2K stage that takes the 768p base and regenerates it at 2K resolution while preserving the original context. On MiniMax3.org, 2K output is available at 12 credits per second.
Flexible Duration (4–15 Seconds)
Generated clips range from 4 to 15 seconds. Duration is selectable per generation. Short clips (4–5s) work well for product B-roll and social media. Longer clips (10–15s) suit cinematic sequences, narrative scenes, and storyboard previews.
Prompt Controllability & Camera Direction
MiniMax H3 responds to detailed camera directions in the text prompt: orbit, push-in, tracking, handheld, static tripod, bullet-time, and more. Combined with image or video references, the model can replicate specific camera movements from reference footage while applying new subjects or environments.
MiniMax H3 Input and Output Modalities
| Input | Output | Workflow | Guide |
|---|---|---|---|
| Text | Video | Text-to-Video | MiniMax H3 Prompt Guide |
| Image (first frame) | Video | Image-to-Video | MiniMax H3 Image to Video |
| Image (first + last frame) | Video | FL2VA (First-Last-Frame-to-Video) | MiniMax H3 Guide |
| Image + Video + Audio | Video | Reference-to-Video (Ref2VA) | MiniMax H3 Reference Video |
| Video + Text | Video | Video-to-Video | MiniMax H3 Video to Video |
| Audio + Text | Video + Audio | Audio-synced generation | MiniMax H3 Guide |
Use Image-to-Video when an image must anchor the opening or ending frame. Use Video-to-Video when an existing source clip is the main object you want to transform. Use Reference-to-Video when multiple assets each need separate roles such as identity, motion, camera, or voice.
MiniMax H3 Resolution and Duration
| Resolution | Description | Availability | Cost |
|---|---|---|---|
| 768p | Standard base resolution | H3-Base (open weights) | 6 credits/second |
| 2K | Full-quality output | Complete H3 system (H3-Regenerate-2K stage) | 12 credits/second |
All resolutions support the full 4–15 second duration range. Generation time scales with resolution and duration. For hardware requirements when running H3-Base locally, see the MiniMax H3 ComfyUI Guide.
MiniMax H3 vs Other MiniMax Models
| Model | Type | Duration | Resolution | Audio | Full Comparison |
|---|---|---|---|---|---|
| MiniMax H3 | Multimodal video | 4–15s | 768p / 2K | Native stereo | — |
| MiniMax H3 Turbo | Video (accelerated) | 4–15s | 768p | Yes | H3 Turbo Guide |
| MiniMax M3 | LLM / agent | — | — | — | M3 vs H3 |
| Seedance 2.0 | Video | — | — | — | H3 vs Seedance |
MiniMax H3 is the core video model. Turbo refers to community-built acceleration layers on top of H3-Base. M3 is MiniMax's separate language and coding model family.
How to Access the MiniMax H3 Video Model
Web (MiniMax3.org)
Generate MiniMax H3 videos directly in your browser with the embedded generator on this page. No install required.
MiniMax H3 API
Integrate MiniMax H3 into your pipeline via REST API. Python, cURL, and SDK examples available.
ComfyUI
Run H3-Base locally with ComfyUI node graphs. Supports FL2VA, Ref2VA, LoRA stacking, and quantized weights.
Hugging Face / Open Weights
Download H3-Base weights (FL2VA & Ref2VA) from Hugging Face. Available in FP16, INT8, and GGUF formats.
FL2VA vs Ref2VA: MiniMax H3 Task Families
FL2VA
First-Last-frame-to-Video-and-Audio
FL2VA covers text-to-video and first/last-frame workflows. It generates video from text prompts alone or from a pair of anchor images that define the opening and closing frames. Best for text-driven creative generation and frame-controlled animation.
Ref2VA
Reference-to-Video-and-Audio
Ref2VA handles multimodal reference-guided generation. It takes multiple image, video, and audio references, each assigned a specific role (identity, motion, camera, voice), and generates video that respects all reference constraints simultaneously.
MiniMax H3 Video Model FAQ
What is the MiniMax H3 video model?
MiniMax H3 is a general-purpose multimodal video generation model developed by MiniMax. It accepts text, image, video, and audio references as input and produces video clips up to 15 seconds at 768p or 2K, with optional native synchronized audio.
Who developed MiniMax H3?
MiniMax H3 was developed by MiniMax, a Chinese AI company. It powers their consumer product Hailuo AI and is available as open weights through Hugging Face and GitHub.
What resolutions does MiniMax H3 support?
MiniMax H3 supports 768p output from H3-Base (open weights) and up to 2K through the complete H3 system, which includes the H3-Regenerate-2K upscaling stage.
How long can MiniMax H3 videos be?
MiniMax H3 generates clips between 4 and 15 seconds. Duration is selectable per generation. Longer content requires stitching multiple clips.
Does MiniMax H3 generate audio?
Yes. MiniMax H3 can generate native, synchronized stereo audio alongside video. Audio generation is supported on compatible workflows and can use audio references for voice, music, or sound effects.
Is MiniMax H3 open source?
The H3-Base model weights (FL2VA and Ref2VA task families) are publicly available on Hugging Face and GitHub. The complete H3 system, including H3-Context-IR and H3-Regenerate-2K, is not fully open.
What hardware do I need to run MiniMax H3 locally?
H3-Base FP16 needs at least 24 GB VRAM (RTX 4090, A100). FP8 runs on 16–20 GB cards (RTX 4080). INT8 or GGUF quantized weights extend to 12 GB GPUs. Apple Silicon needs 32 GB+ unified memory.
How is MiniMax H3 different from Hailuo AI?
Hailuo AI is a consumer-facing video generation product built on MiniMax H3. H3 is the underlying model with open weights, API access, and ComfyUI integration. Hailuo provides a simpler web interface aimed at end users.
What is the difference between FL2VA and Ref2VA?
FL2VA (First-Last-frame-to-Video-and-Audio) covers text-to-video and first/last-frame workflows. Ref2VA (Reference-to-Video-and-Audio) handles multimodal reference-guided generation. Both are task families within H3-Base.
How do I try the MiniMax H3 video model?
You can try MiniMax H3 for free on this page using the embedded generator. New users receive 10 credits daily, with up to 30 total free credits. No download or API key required.
Try the MiniMax H3 Video Model
Generate your first MiniMax H3 video in seconds. Text-to-video, image-to-video, and reference-guided workflows — all in your browser.
10 free credits per daily check-in · Up to 30 total
Data References
MiniMax H3 official launch announcement
MiniMax H3 on Hugging Face (open weights)
Published 2026-08-18. Last reviewed 2026-08-18. MiniMax3.org is an independent third-party platform and is not affiliated with, endorsed by, or operated by MiniMax. Model specifications are sourced from official MiniMax documentation and the Hugging Face model card.