MiniMax H3 Complete Guide: Architecture, Open-Weight License, Benchmarks vs Gemini Omni, Seedance 2.0, Kling 3.0, LTX-2.3 & More

By AiCybr Editorial · Published August 5, 2026

What This Article Covers

MiniMax H3 is an omni-modal generative video system that can accept combinations of text, images, video, and audio, then generate video with synchronized native stereo audio. The hosted workflow supports clips from 4 to 15 seconds, output at 24 FPS, and resolutions up to 2K.

H3 is technically important for two reasons:

1. It is the first publicly downloadable audio-video model to perform near the top proprietary systems in current blind preference testing.
2. It is not a conventional text-to-video checkpoint. The complete product combines multimodal reasoning, reference processing, joint audio-video generation, and a separate 2K regeneration stage.

This guide covers:

• H3’s complete architecture and the parts that are actually downloadable
• The difference between open weights and open source
• Independent benchmark comparisons against open and proprietary models
• Text-to-video, image-to-video, and video-editing results with audio
• API pricing and realistic local hardware requirements
• Comparisons with Gemini Omni Flash, Seedance 2.0, Kling 3.0, Veo 3.1, Runway Aleph 2.0, LTX-2.3, and Wan 2.2
• A practical decision framework for creators, developers, and self-hosters

Important: H3’s leaderboard position is promising but very recent. Arena ratings change as more votes arrive, and the confidence intervals of H3 and the leading proprietary models overlap in several categories. Treat the results as a current snapshot rather than a permanent ranking.

MiniMax H3 at a Glance

Sources: MiniMax H3 model card, MiniMax H3 licence, and MiniMax API pricing.

Is MiniMax H3 Really Open Source?

The most accurate description is:

MiniMax H3 is an open-weight model with available inference code, not a fully open-source system under a permissive licence.

There are three separate issues.

1. Only H3-Base Is Fully Downloadable

The official production workflow has three major stages:

A local installation can run the core H3-Base model. Reproducing MiniMax’s complete official 2K workflow still requires hosted components unless a developer builds substitutes for the missing orchestration and upscaling/regeneration stages.

2. The Licence Is Restrictive

The MiniMax H3 Community License is not equivalent to Apache 2.0, MIT, or another standard open-source licence.

Key restrictions include:

• The licence excludes the European Union, United Kingdom, Republic of Korea, and United States from its applicable territory.
• Commercial products and services earning more than $20 million in annual revenue require separate written authorization.
• Commercial products using H3 must prominently display “MiniMax H3” in their user interface.
• H3, its derivatives, outputs, or results may not be used to improve another AI model, except H3 or its derivatives.
• Hosted services must implement and maintain safeguards against prohibited uses.
• MiniMax states that it claims no rights over generated outputs, but users remain responsible for those outputs.

The embedded Qwen3-VL-32B encoder is separately licensed under Apache 2.0, but that does not convert the overall H3 package into an Apache-licensed model.

3. “Open Weights” Still Matters

Despite the licence and hosted-component limitations, releasing the generative base model is substantial. Developers can:

• Inspect and modify the inference implementation
• Run the core model on controlled infrastructure
• Fine-tune the complete released weights
• Build custom preprocessing and reference-conditioning workflows
• Avoid sending source media to a third-party generation API where the licence permits deployment

The practical verdict is therefore neither “fully open source” nor “API-only.” H3 sits between those categories.

System Architecture

1. H3-Context-IR: Multimodal Planning and Orchestration

H3-Context-IR is the front end of the official workflow. It interprets relationships among text prompts, images, reference videos, and reference audio. Its tasks include:

• Instruction parsing
• Cross-modal association
• Temporal understanding
• Logical reasoning across references
• Resolving underspecified details
• Serializing the result into a structured representation accepted by H3-Base

MiniMax does not release this component because it relies on multiple hosted models and services. The company provides an API and prompting guidance so developers can reproduce part of its behavior.

This distinction matters in practice: downloading H3-Base does not automatically reproduce every result shown in the hosted MiniMax product.

2. Multimodal Encoder: Qwen3-VL-32B

H3 uses the full Qwen3-VL-32B model as its context encoder rather than a small text-only encoder.

The H3 model card states that it extracts hidden states from layer 50 of Qwen3-VL. This allows the generative model to condition on richer representations of:

• Natural-language instructions
• Image identity and composition
• Video motion and temporal cues
• Audio characteristics
• Relationships among multiple references

The “33B” label commonly associated with H3 refers to the H3-Omni-Transformer. The complete inference pipeline also includes the Qwen3-VL encoder and the visual and audio autoencoders, so its total memory footprint is much larger than that of a typical standalone 33B language model.

3. H3-VisualVAE

The visual autoencoder compresses raw video before the transformer processes it.

The high compression ratio reduces the sequence length passed into the transformer. That is essential because even a short 768p video contains a very large number of pixels and frames.

4. H3-AudioVAE

H3 generates audio jointly with video rather than adding a separate sound-effects pass after video generation.

The audio autoencoder:

• Processes left and right channels independently using shared encoder and decoder weights
• Recombines the decoded channels into stereo output
• Compresses 32 kHz audio to latent tokens at a rate of 40 Hz
• Is trained to preserve audio reconstruction quality while producing a latent space that the generative transformer can learn

This design supports synchronized dialogue, ambience, music, and physical sound effects within the same denoising process as the video.

5. H3-Omni-Transformer

The core generator is a 33B-parameter dense, single-stream transformer.

MiniMax notes that approximately 13B parameters reside in AdaLN-related branches. Their modulation outputs can be precomputed and cached, so inference-only deployments do not necessarily need to keep all of those parameters loaded continuously.

6. H3-Regenerate-2K

The hosted 2K stage does more than perform a conventional pixel-space upscale. It feeds the generated 768p result and the original multimodal context back into H3, then regenerates a higher-resolution result.

Potential advantages over a simple super-resolution model include:

• Recovering details using the original prompt and references
• Correcting visual information rather than merely sharpening it
• Preserving semantic context during high-resolution regeneration

The trade-off is that this stage is not part of the downloadable release.

Input Modes and Capabilities

Reference Limits

The official model card lists the following H3-Base-Ref2VA limits:

• Up to 9 images
• Up to 3 video clips
• Each video reference: 2–15 seconds
• Total reference-video duration: 15 seconds
• Up to 3 audio clips
• Each audio reference: 2–15 seconds
• Total reference-audio duration: 15 seconds
• Audio references must accompany an image or video reference
• Up to 12 total files

Output Specifications

• Duration: 4–15 seconds
• Frame rate: 24 FPS
• Audio: 32 kHz stereo
• Default short edge: 768 pixels
• Hosted maximum: 2K
• Supported aspect ratios include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16

Stable dialogue support is documented for Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

Independent Benchmark Comparison

Methodology

MiniMax does not currently publish a comprehensive, independently replicated VBench-style table for H3. To avoid mixing incompatible vendor evaluations, the comparisons below use the Artificial Analysis Video Arena.

The arena:

• Shows users two outputs generated from the same prompt
• Hides model identities during voting
• Converts pairwise preferences into Elo ratings
• Publishes 95% confidence intervals and sample counts
• Separates text-to-video, image-to-video, and video-editing categories
• Separates models that generate audio from those that do not

An Elo difference is not a percentage-point quality difference. Scores also cannot be compared directly across different arena categories.

Snapshot date: August 5, 2026.

Text-to-Video With Audio

*Artificial Analysis normalizes this field as the cost to generate one minute on the creator’s API at default settings. Resolution, duration options, and billing rules still differ among providers.

Interpretation

• H3 is second overall and first among downloadable open-weight models.
• Gemini Omni Flash leads by only 6 Elo points.
• Their uncertainty ranges overlap: Gemini spans approximately 1,237–1,251, while H3 spans approximately 1,229–1,247.
• It is therefore not statistically defensible to claim that Gemini is conclusively better based on this snapshot alone.
• H3 leads Seedance 2.0 by 15 Elo points, although Seedance has substantially more votes.
• H3’s lead over LTX-2.3 is very large in this arena, but LTX is far cheaper and substantially easier to adapt for local workflows.

Source: Artificial Analysis Text-to-Video Leaderboard.

Image-to-Video With Audio

Interpretation

• Seedance 2.0 is nominally first, Gemini Omni Flash second, and H3 third.
• All three occupy overlapping statistical rank ranges.
• H3 trails Seedance by only 9 Elo points and Gemini by 7.
• H3 is again the highest-ranked downloadable open-weight model with audio.
• The large gap between H3 and LTX-2.3 suggests that H3’s heavy architecture currently buys meaningful perceptual quality, not merely a larger parameter count.

Source: Artificial Analysis Image-to-Video Leaderboard.

Video Editing With Audio

Interpretation

• H3 currently ranks first in video editing with audio.
• Its 8-point lead over Gemini Omni Flash is smaller than the combined uncertainty range, so the result should be described as a current nominal lead, not a decisive victory.
• H3’s stronger separation from Seedance, Aleph 2.0, and Kling 3.0 Omni is more meaningful within this specific arena.
• Aleph 2.0 remains specialized for controlled edits to existing footage and supports longer clips than H3, so its lower arena rating does not make it inferior for every editing workflow.

Source: Artificial Analysis Video Editing Leaderboard.

Open-Weight Audio-Video Models Only

H3 is currently the clear open-weight leader in these two audio-enabled arenas. However, this table does not measure:

• Fine-tuning ease
• Minimum hardware
• Generation speed on local hardware
• Licence suitability
• Training transparency
• Reproducibility of the hosted product
• Long-duration production workflows

LTX-2.3 remains the more practical model for many local developers despite its lower blind preference score.

Why You Should Ignore Unverified “VBench 84.7” Claims

Several secondary articles and social posts attach precise VBench-style scores to H3 without linking a MiniMax technical report, reproducible evaluation configuration, or official model card table.

As of this research date:

• MiniMax’s H3 model card does not provide a full public VBench comparison table.
• The public release does not contain an independently reproduced quality report covering all major competitors.
• Different VBench variants, prompt sets, sampling settings, and aggregation methods can produce materially different scores.

A benchmark number should not be presented as verified unless its source and evaluation configuration are available. The independent arena data above is currently the most defensible cross-model comparison.

H3 Compared With Similar Models

Closed-Source and Hosted Models

Gemini Omni Flash

Google’s Gemini Omni Flash accepts combinations of text, images, audio, and video, then generates or conversationally edits video with native audio.

Key characteristics:

• 10-second video generation
• Native audio
• Up to five photo references in the consumer product
• Video-to-video editing
• Multi-turn conversational editing
• Strongest current text-to-video arena score
• Lower listed API cost than H3 in Artificial Analysis
• Closed weights

Best fit: users who prioritize an integrated hosted experience, conversational iteration, and strong quality without managing infrastructure.

Dreamina Seedance 2.0

ByteDance’s Seedance 2.0 is a native multimodal audio-video system described in an official model card paper.

Key characteristics:

• Text, image, video, and audio inputs
• Native synchronized audio-video output
• 4–15 second clips
• 480p and 720p output in the documented release
• Up to 3 video, 9 image, and 3 audio references
• Current nominal leader for image-to-video with audio
• Closed weights

Best fit: reference-heavy image-to-video generation and users already working in ByteDance’s creative ecosystem.

Kling 3.0 and Kling 3.0 Omni

Kuaishou describes Kling 3.0 as an all-in-one multimodal family supporting text, images, audio, and video.

Key characteristics:

• Up to 15-second video
• Native audio in multiple languages, dialects, and accents
• Text-to-video, image-to-video, reference-to-video, and in-video editing
• Multi-shot storytelling and storyboard control
• Strong reference consistency
• Closed weights

Best fit: cinematic multi-shot generation, multilingual dialogue, and controlled storytelling through a hosted product.

Veo 3.1

Veo 3.1 remains a strong proprietary audiovisual generator, although Google states that Gemini Omni is replacing Veo in the Gemini application.

Key characteristics:

• Native audio generation
• Mature Google ecosystem integration
• Strong image-to-video performance
• Significantly higher listed cost than H3 in the current arena snapshot
• Closed weights

Best fit: existing Google media workflows or applications that already depend on Veo APIs and tooling.

Runway Aleph 2.0

Aleph 2.0 is not primarily a from-scratch audiovisual generator. It is a specialized in-context editing model.

Key characteristics:

• Edit one key frame and propagate the change through the clip
• Preserve unrequested elements
• Apply edits across multiple shots
• Up to 30 seconds of 1080p source video
• Image preview before committing to the video edit
• Closed weights

Best fit: advertising, product variation, relighting, wardrobe or background replacement, and controlled modification of real footage.

Open-Weight Alternatives

LTX-2.3

LTX-2.3 is H3’s most direct downloadable competitor because it also generates synchronized audio and video in one model.

Strengths:

• Lower hardware and serving barrier than H3
• Fast distilled checkpoints
• Complete local creative pipeline
• Trainable base and published adaptation tooling
• Much lower creator-API cost in the current arena

Limitations:

• Substantially lower blind preference scores than H3 for current audio-enabled T2V and I2V
• Custom licence rather than Apache/MIT
• Audio quality can be weaker for non-speech generation according to its model card
• Prompting style strongly affects results

Best fit: local experimentation, fine-tuning, cost-sensitive generation, and developers who value practical control more than maximum current arena quality.

Wan 2.2

Wan 2.2 is not a direct H3 substitute for joint native audio-video generation, but it remains a useful open-weight video family.

Available variants include:

• 14B MoE text-to-video
• 14B MoE image-to-video
• 5B combined text/image-to-video
• 720p generation at 24 FPS
• A 5B variant designed to run on a consumer RTX 4090
• Larger 14B variants that require much more memory

Best fit: open video generation without a requirement for native jointly generated stereo audio.

Capability Comparison

Pricing

Official MiniMax H3 API Pricing

Equivalent output-only cost:

Reference-video charges and extra image charges can increase the total.

Arena-Listed Creator API Cost

This is useful for directional comparison, not precise procurement. Providers use different default resolutions, clip durations, quality tiers, reference-input charges, and retry policies.

Local Deployment

Supported Checkpoints

The open repository provides two BF16, CFG-distilled families:

1. FL2VA for text-to-audio-video and first/last-frame conditioning
2. Ref2VA for image, video, and audio reference conditioning

Each checkpoint includes:

<TASK>/
├── model_index.json
├── processor/
├── tokenizer/
├── text_encoder/
├── transformer/
├── visual_vae/
└── audio_vae/

Download a Single Task Family

hf download MiniMaxAI/MiniMax-H3 
  --include "model_index.json" "modular_model_index.json" "FL2VA/*" 
  --local-dir MiniMax-H3

Download only the task family required by your workload. Pulling both FL2VA and Ref2VA, plus duplicate framework layouts, consumes much more storage.

Start an SGLang Server

A verified four-H200 configuration:

sglang serve 
  --model-path MiniMaxAI/MiniMax-H3 
  --model-variant fl2va 
  --num-gpus 4 
  --ulysses-degree 4 
  --performance-mode speed 
  --port 30010

A verified four-H100 configuration:

sglang serve 
  --model-path MiniMaxAI/MiniMax-H3 
  --model-variant fl2va 
  --num-gpus 4 
  --tp-size 2 
  --ulysses-degree 2 
  --performance-mode speed 
  --port 30010

Submit a Local Text-to-Audio-Video Job

video_id=$(
  curl -sS -X POST http://127.0.0.1:30010/v1/videos 
    -H "Content-Type: application/json" 
    -d '{
      "model": "MiniMaxAI/MiniMax-H3",
      "prompt": "A cinematic rain-soaked street at night. A bicycle passes the camera while distant traffic, rain, and a soft bell are heard.",
      "seconds": 5,
      "task": "t2va",
      "conditions": [],
      "target": {
        "short_edge": 768,
        "aspect_ratio": "16:9",
        "duration_seconds": 5.0
      },
      "num_outputs_per_prompt": 1,
      "num_inference_steps": 50,
      "flow_shift": 12.0,
      "audio_flow_shift": 3.0,
      "seed": 1101
    }' | jq -r '.id'
)

while true; do
  status=$(curl -sS "http://127.0.0.1:30010/v1/videos/${video_id}" | jq -r '.status')
  [ "$status" = "completed" ] && break
  [ "$status" = "failed" ] && exit 1
  sleep 1
done

curl -sS -L 
  "http://127.0.0.1:30010/v1/videos/${video_id}/content" 
  -o minimax-h3-output.mp4

The documented local output contract is an MP4 with H.264 video at 24 FPS and one AAC stereo track at 32 kHz.

Realistic Hardware Requirements

H3 is downloadable, but it is not a normal single-GPU desktop model.

Verified SGLang Results

The following figures come from SGLang’s H3 deployment documentation. Workloads and optimizations differ, so do not compare every row as a universal GPU ranking.

The verified dual-RTX-5090 system also used a 377 GiB host, and SGLang recommends a 384 GiB-class machine for that profile.

Practical Meaning

• A single 24GB or 32GB consumer GPU is not a realistic target for the official full checkpoint.
• Dual RTX 5090 deployment is possible only with aggressive layerwise offload and very large system memory.
• The 50-step consumer result took more than nine minutes to produce five seconds of video.
• Datacenter GPUs provide far better latency, but hardware cost is substantial.
• FP8 reduces memory significantly on supported hardware, with only modest latency improvement in the published B300 test.
• API use is economically rational for low or moderate volume.
• Self-hosting becomes attractive mainly for privacy, customization, high sustained utilization, or infrastructure research.

Real-World Strengths

1. Frontier-Level Open-Weight Quality

H3 is the only current downloadable model positioned inside the top statistical group in all three Artificial Analysis audio-enabled categories covered here.

2. Native Joint Audio and Video

The model does not require a separate speech, Foley, music, or lip-sync pipeline for basic generation. Joint prediction can improve temporal coherence between sound and visible events.

3. Rich Multimodal Conditioning

The Ref2VA checkpoint supports images, video, and audio references in the same request. This is useful for:

• Character identity
• Product appearance
• Camera movement
• Acting and motion
• Voice or sound character
• Style and scene continuity

4. First-and-Last-Frame Control

FL2VA can preserve an exact opening frame, ending frame, or both. That provides more deterministic shot construction than a general reference-image workflow.

5. Full Released Weights for Fine-Tuning

MiniMax releases the complete H3-Base weights rather than only an API or a small distilled derivative.

6. Competitive API Pricing

At $7.80 per output minute for 2K, H3 costs more than Gemini Omni Flash in the current comparison but less than Seedance 2.0, Kling 3.0 Pro, and Veo 3.1.

Limitations and Risks

1. The Complete Product Is Not Open

H3-Context-IR and H3-Regenerate-2K remain hosted. The downloadable model produces the 768-pixel-short-edge base result, not the complete official 2K system.

2. The Licence Excludes Major Markets

The territorial exclusions cover the United States, European Union, United Kingdom, and Republic of Korea. This can make the checkpoint unusable for global products without a separate licence.

3. Local Hardware Requirements Are Extreme

The dual-RTX-5090 profile requires approximately 384GB of host memory and remains slow at the full 50-step setting.

4. Results Are Early

H3 entered the arena recently and has fewer samples than several established competitors. Its Elo rating may move meaningfully as the vote count grows.

5. Arena Scores Do Not Measure Every Production Requirement

They do not directly measure:

• Prompt safety
• Copyright or identity risk
• Temporal reliability across repeated generations
• API uptime
• Batch throughput
• Fine-tuning stability
• Exact character consistency across many independent clips
• Cost of rejected generations
• Moderation false positives
• Long-form sequence continuity

6. 15 Seconds Is Still Short

H3 can generate longer clips than several older video models, but a professional sequence still requires shot planning, multiple generations, continuity management, editing, and post-production.

7. Sparse-Attention Serving Is Not Yet the Default Open Path

MiniMax reports using sparse attention during training, while the current open release and common serving path primarily use full attention. Future optimized kernels may improve deployment efficiency.

8. Output Responsibility Remains With the User

The licence places responsibility for outputs and downstream use on the user or service operator. Production deployments need consent, rights management, content safeguards, and clear AI-generation labelling policies.

Head-to-Head Decisions

MiniMax H3 vs Gemini Omni Flash

Verdict: Gemini is the easier hosted product; H3 is the stronger choice when local control or model modification is essential and the licence permits use.

MiniMax H3 vs Seedance 2.0

Verdict: H3 is the more flexible developer platform and stronger all-round arena performer. Seedance remains highly competitive for image-conditioned generation.

MiniMax H3 vs Kling 3.0

Verdict: H3 is stronger for technical ownership, cost, and current blind preference. Kling remains attractive for polished hosted storytelling controls.

MiniMax H3 vs LTX-2.3

Verdict: H3 is the quality-first choice. LTX-2.3 is the practical local-development choice.

MiniMax H3 vs Runway Aleph 2.0

Verdict: Use H3 for generative audiovisual creation and flexible references. Use Aleph when preserving an existing shot while making a targeted production edit matters more than generative range.

Decision Framework

Choose MiniMax H3 if:

• You need the highest-current-quality downloadable audio-video model
• You require text, image, video, and audio references in one workflow
• You need native stereo audio rather than a separate sound pipeline
• You want to fine-tune the released base weights
• You can operate within the licence’s territorial and commercial restrictions
• You have datacenter-class hardware or prefer using the MiniMax API
• First-and-last-frame control is important
• You are building research or production systems around multimodal video generation

Choose Gemini Omni Flash if:

• You want the simplest high-quality hosted workflow
• Conversational, multi-turn editing is central
• Ten-second clips are sufficient
• You do not need local weights
• Lower listed API cost matters
• Google ecosystem integration is valuable

Choose Seedance 2.0 if:

• Image-to-video quality is your primary criterion
• Rich multimodal references are important
• A hosted ByteDance workflow is acceptable
• 720p documented output is sufficient for the base generation stage

Choose Kling 3.0 if:

• You prioritize multilingual spoken scenes
• Multi-shot cinematic storytelling and storyboard controls matter
• You prefer a polished hosted creator tool
• Local deployment is not required

Choose LTX-2.3 if:

• You need a more manageable local model
• Fast distilled inference matters more than maximum quality
• You plan to train LoRAs or modify the pipeline
• Budget is the primary constraint
• A custom community licence is acceptable

Choose Runway Aleph 2.0 if:

• You are editing existing footage rather than generating from scratch
• Precise preservation of unchanged areas matters
• You need clips up to 30 seconds
• Previewing an edit before rendering is valuable

Choose Wan 2.2 if:

• You want an established open-weight video model
• Native audio generation is not required
• A 5B consumer-GPU variant is more useful than frontier output quality
• You need a broad open ecosystem around T2V and I2V

Production Recommendations

1. Validate the Licence Before Technical Integration

Do not build a product around H3 and review the licence later. Confirm:

• Deployment territory
• Annual product/service revenue
• Required UI attribution
• User terms and acceptable-use controls
• Whether generated outputs may enter training pipelines
• Distribution obligations
• Rights to source images, video, voices, music, and likenesses

2. Benchmark Your Actual Workflow

A public arena cannot answer whether H3 works for a particular product. Build a fixed internal test set covering:

• Human faces and hands
• Product logos and text
• Dialogue in required languages
• Camera movement
• Fast physical motion
• Multi-character scenes
• Reference identity
• Voice consistency
• Music and ambience
• Editing preservation
• Failure recovery and retries

Run blind reviews and track success rate, not only the best sample.

3. Separate Generation Cost From Usable-Shot Cost

A model that costs $0.65 for a five-second 2K output may require multiple attempts. Track:

effective cost per usable shot
= total generation spend / accepted outputs

A lower nominal API price can be more expensive if rejection rates are high.

4. Start With the API

For most teams, the sensible sequence is:

1. Test the hosted 2K workflow
2. Establish quality and retry rates
3. Compare against Gemini, Seedance, Kling, and LTX on the same prompts
4. Estimate monthly generation volume
5. Evaluate whether privacy or customization justifies self-hosting
6. Move to H3-Base locally only after infrastructure economics are clear

5. Treat Local 768p and Hosted 2K as Separate Products

Do not assume that a locally generated 768p result plus a generic upscaler will match MiniMax’s hosted regenerated 2K output. Evaluate them independently.

6. Keep Human Review in the Publishing Pipeline

Review generated content for:

• Identity misuse
• Copyrighted characters or brands
• Misleading realism
• Dialogue errors
• Accidental text or signage
• Audio artefacts
• Cultural and language issues
• Platform labelling requirements

Final Verdict

MiniMax H3 is one of the most consequential open-weight video releases of 2026. Its significance is not that it is universally better than every closed model. The stronger claim supported by current evidence is narrower:

H3 brings downloadable weights into the same statistical quality group as the best hosted audiovisual generators in current blind preference testing.

It ranks:

• Second for text-to-video with audio
• Third for image-to-video with audio
• First nominally for video editing with audio
• First among open-weight models in both current audio-enabled generation arenas

Those results are unusually strong for a downloadable checkpoint. They also come with major qualifications:

• Confidence intervals overlap the category leaders
• The complete 2K workflow is not open
• The licence excludes several major markets
• Local inference requires extreme hardware
• The model is too new for long-term production reliability conclusions

For most users, the hosted H3 API is the practical starting point. For researchers and advanced infrastructure teams, H3-Base is a rare opportunity to inspect, adapt, and self-host a frontier-class joint audio-video generator. For ordinary local creators, LTX-2.3 remains much easier to run, while Gemini Omni, Seedance, Kling, and Runway provide more accessible hosted workflows.

The correct choice depends less on a single leaderboard rank and more on five factors: licence, reference-control needs, acceptable infrastructure cost, editing workflow, and cost per usable output.

References

MiniMax H3

• MiniMax H3 official Hugging Face model card
• MiniMax H3 Community License Agreement
• MiniMax API pay-as-you-go pricing
• SGLang MiniMax H3 deployment guide and benchmarks

Independent Benchmarks

• Artificial Analysis Text-to-Video Leaderboard
• Artificial Analysis Image-to-Video Leaderboard
• Artificial Analysis Video Editing Leaderboard

Comparable Models

• Google: Introducing Gemini Omni
• Gemini Omni product overview
• Seedance 2.0 model card paper
• Kuaishou: Kling AI 3.0 launch
• Runway Aleph 2.0
• LTX-2.3 official model card
• Wan 2.2 T2V-A14B official model card

Last updated: August 5, 2026. Leaderboard scores are dynamic and should be rechecked before publication or procurement decisions.