Quick Answer: MiniMax H3 is the better starting point for a hosted workflow with 5 to 15 second native 2K output, multimodal references, and no local GPU setup. Wan 2.2 fits teams that want Apache 2.0 weights, ComfyUI or Diffusers integration, and control over their inference stack. Both have downloadable weights. MiniMax H3 offers a hosted 2K route, while its public local base currently stops at 768p and uses separate MiniMax APIs for full 2K regeneration. Wan 2.2 publishes several task-specific checkpoints, including a 5B text and image model that can run with 24 GB of VRAM. Choose by deployment, audio, resolution, and the price of the route you will use. No controlled paid test was run, so this article does not claim a visual-quality winner.
MiniMax H3 model page. Source: https://minimaxh3.tv/models/minimax-h3. Captured August 19, 2026; region: United States; product UI evidence.
MiniMax H3 vs Wan 2.2 at a glance
| Decision factor | MiniMax H3 | Wan 2.2 |
|---|---|---|
| Best starting point | Hosted creation, native 2K, mixed references | Local deployment, model customization, open tooling |
| Current output evidence | Official system supports 4 to 15 seconds, up to 2K, 24 FPS, stereo audio; minimaxh3.tv exposes 5 to 15 seconds at 2K | Official T2V and I2V A14B checkpoints support 480p and 720p; TI2V-5B supports 720p at 24 FPS |
| Inputs | Text, first and last frames, images, video, and audio references | Separate T2V, I2V, TI2V, S2V, and Animate checkpoints |
| Local deployment | H3-Base weights are public; complete 2K regeneration still uses an official API module | Apache 2.0 weights and inference code are public |
| Local hardware signal | Official open-source page documents local H3-Base, but does not present it as a consumer-laptop workflow | A14B examples require at least 80 GB VRAM; TI2V-5B can run with at least 24 GB VRAM |
| Audio | Official H3 system generates stereo audio; this site's current workflow accepts audio references but does not generate audio | Standard T2V and I2V are video models; S2V is a separate speech-to-video route |
| Price evidence | minimaxh3.tv publishes current USD plan and estimated per-second retail rates | Alibaba Cloud publishes Wan 2.2 per-second API prices by region and resolution |
Checked August 19, 2026. This is not a same-prompt benchmark.
How this comparison was made
The research used MiniMax's launch and open-source pages, the current minimaxh3.tv interfaces, the official Wan 2.2 repository and model card, and Alibaba Cloud pricing. Current results for MiniMax H3 vs Wan 2.2 mix model pages, technical documentation, videos, and comparisons. The query is a consideration-stage choice between a managed workflow and a deployable model stack.
No new generation was purchased. Vendor demos can prove that a workflow exists, but not that it will win on your assets and acceptance criteria.
Current minimaxh3.tv generator. Source: https://minimaxh3.tv/ai-video-generator/text-to-video. Captured August 19, 2026; region: United States. The page shows available controls, not a benchmark result.
MiniMax H3 is a managed workflow first
MiniMax describes H3 as a general-purpose multimodal system. Its launch page says the model understands text, images, video, and audio in one context and generates stereo audio with video up to 15 seconds and 2K. The public specification lists 4 to 15 seconds, 24 FPS, and 32 kHz stereo audio.
On minimaxh3.tv, H3 is exposed as three practical modes: text, start or end frame, and multimodal reference generation. The current interface accepts up to nine images, three video clips, and three audio clips, with no more than 12 reference files in total. It outputs 5 to 15 second 2K clips. The site accepts audio as a reference, but it does not currently generate audio. That product boundary matters if your delivery needs spoken dialogue or a finished soundtrack.
MiniMax H3 launch page. Source: https://www.minimax.io/blog/minimax-h3. Captured August 19, 2026. Official model claims only.
The hosted route removes model downloads and GPU provisioning from the first decision. The free account tier includes no generation credits, so "start free" means inspecting the workflow, not receiving a guaranteed free video.
MiniMax H3 is open, with an important 2K boundary
MiniMax released H3-Base checkpoints on August 3, 2026. One handles text or first and last frame audio-video generation; the other handles multimodal reference audio-video generation. The local base produces 768p output.
The complete 2K system uses H3-Context-IR, H3-Base, and H3-Regenerate-2K. MiniMax says the regeneration module is not yet open-sourced. Its full 2K workflow combines local H3-Base with official APIs for context processing and regeneration. A downloaded checkpoint does not reproduce the whole hosted 2K route by itself.
MiniMax H3 open-source announcement. Source: https://www.minimax.io/news/minimax-h3-open-source. Captured August 19, 2026. The page distinguishes the local base from the complete 2K workflow.
H3 therefore offers hosted access and public base weights, with a hybrid dependency for full-resolution reproduction.
Wan 2.2 gives you a broader open checkpoint family
Wan 2.2 publishes inference code and multiple checkpoints through the official Wan-Video repository. The main families include T2V-A14B for text to video, I2V-A14B for image to video, TI2V-5B for both text and image input, S2V-14B for speech-driven video, and Animate-14B for character animation and replacement.
Wan-Video/Wan2.2 official repository. Source: https://github.com/Wan-Video/Wan2.2. Captured August 19, 2026. Repository status and public project activity are shown as factual context.
The repository uses Apache 2.0 and links official Hugging Face and ModelScope downloads. It fits teams that need to modify inference, integrate ComfyUI or Diffusers, or keep assets on their own infrastructure.
Official Wan 2.2 model table. Source: https://github.com/Wan-Video/Wan2.2#model-download. Captured August 19, 2026. The table records model families and documented resolutions.
You must still choose a checkpoint, install dependencies, manage GPU memory, monitor jobs, and store outputs. Local inference removes per-generation API billing, not hardware and maintenance costs.
Hardware and speed change the Wan decision
The repository says the single-GPU 720p A14B examples require at least 80 GB of VRAM. TI2V-5B supports 720p at 24 FPS and can run with at least 24 GB when memory-saving options are enabled. It reports a five-second 720p result in under nine minutes on one consumer GPU without specific optimization.
Official Wan 2.2 run instructions. Source: https://github.com/Wan-Video/Wan2.2#run-wan22. Captured August 19, 2026. Hardware figures are vendor documentation, not measurements from this article.
A14B and 5B are different operating commitments. The 5B model may fit an existing GPU pipeline; a hosted API may cost less for occasional clips.
Output, audio, and control
MiniMax H3's hosted specification reaches 2K and 15 seconds with native stereo audio in the official system. On this site, mixed references are available, while audio generation remains disabled.
Wan 2.2 is more modular. Standard text and image checkpoints focus on video; audio-driven work uses the separate S2V checkpoint with an input audio file or optional CosyVoice support.
Wan's open checkpoints document 480p and 720p. Alibaba Cloud's hosted wan2.2-t2v-plus includes 480p and 1080p. H3's hosted path reaches 2K, while local H3-Base defaults to a 768-pixel short side. These are different deployment routes.
Wan-AI/Wan2.2-T2V-A14B official model card. Source: https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B. Captured August 19, 2026. This confirms the public weight distribution route.
Pricing: compare the route, region, and resolution
The current minimaxh3.tv annual view lists Lite at $9.90 per month billed annually, Pro at $24.90, and Premium at $49.90. It estimates H3 retail cost at about $0.59, $0.46, and $0.42 per generated second. Credits are issued monthly and unused subscription credits reset. These are minimaxh3.tv retail prices, not MiniMax's universal API rate.
minimaxh3.tv pricing. Source: https://minimaxh3.tv/pricing. Captured August 19, 2026; region: United States; currency: USD; annual billing view with monthly credit issuance.
Alibaba Cloud Model Studio lists wan2.2-t2v-plus in China (Beijing) at CNY 0.14 per second for 480p and CNY 0.70 per second for 1080p, with a 50-second quota subject to its stated 90-day eligibility window. The Singapore table lists CNY 0.146785 and CNY 0.733924 per second and states that the free quota is not available outside China (Beijing). Prices and model availability vary by region, so do not convert one table into a global promise.
Alibaba Cloud Model Studio pricing for wan2.2-t2v-plus. Source: https://help.aliyun.com/en/model-studio/model-pricing. Captured August 19, 2026; displayed region: China (Beijing); currency: CNY; billing unit: output second.
For local Wan, add GPU, setup, failed runs, storage, and monitoring. For local H3, include official API steps when the deliverable needs the full 2K workflow. Compare cost per approved clip.
Which one should you use?
Choose MiniMax H3 for browser-based 5 to 15 second 2K video, mixed references, and no model infrastructure.
Choose Wan 2.2 for Apache 2.0 weights, local control, custom nodes, or code-level integration. The 5B model is the practical 24 GB entry point; A14B demands much more memory.
Choose neither on a vendor demo alone. For a quality decision, run the same prompt, input image, aspect ratio, duration, resolution target, and number of attempts. Score action order, subject consistency, object count, camera motion, text rendering, audio behavior, retry count, and usable seconds. This brief did not authorize that paid test, so no quality ranking appears here.
Alibaba Cloud's official Wan 2.2 launch video had 31,203 views and 676 likes when checked on August 19, 2026, passing this article's popularity threshold. It remains a vendor example, not proof that Wan wins a task.
Original source: https://www.youtube.com/watch?v=ktDogWm7Hac. Published by Alibaba Cloud on July 28, 2025. Metrics captured August 19, 2026: 31,203 views and 676 likes.
FAQ
Is MiniMax H3 better than Wan 2.2?
Not universally. H3 has the easier hosted 2K and multimodal workflow. Wan 2.2 offers a more conventional Apache 2.0 local deployment path and broader open tooling. Visual quality needs a controlled same-input test.
Is Wan 2.2 free?
The official weights are available under Apache 2.0, but running them still costs hardware and engineering time. Alibaba Cloud lists a limited free quota only for eligible China (Beijing) users. Other regions have paid per-second pricing.
Can Wan 2.2 run on a 24 GB GPU?
The official repository says TI2V-5B can run with at least 24 GB of VRAM when memory-saving options are used. The A14B single-GPU examples require at least 80 GB.
Does MiniMax H3 run locally?
H3-Base checkpoints can run locally and produce 768p output. MiniMax's documented full 2K reproduction combines the local base with official context and regeneration APIs because the 2K regeneration module is not yet open-sourced.
Which model supports audio?
The official H3 system generates stereo audio. Wan 2.2 uses a separate S2V speech-to-video checkpoint for audio-driven work. The current minimaxh3.tv workflow accepts audio references but does not generate audio.
Which model is cheaper?
There is no route-independent answer. Compare the exact hosted plan, API region, resolution, retries, and local GPU costs. The published prices in this article use different currencies and product scopes and should not be treated as a direct universal rate comparison.
Is minimaxh3.tv operated by MiniMax?
No. minimaxh3.tv is an independent service that provides access to MiniMax H3 and other supported models. It is not affiliated with, endorsed by, or operated by MiniMax or Alibaba Cloud.
Related MiniMax H3 guides
- MiniMax H3 Release Date
- How to Use MiniMax H3
- MiniMax H3 API Guide
- MiniMax H3 Open Source Status
- MiniMax H3 vs Seedance 2.5
- MiniMax H3 vs Veo 3
Open the MiniMax H3 model page to inspect the current hosted workflow, or use the official Wan 2.2 repository to size a local deployment before committing GPU budget.
Sources and methodology
- MiniMax, "MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities": https://www.minimax.io/blog/minimax-h3. Accessed August 19, 2026.
- MiniMax, "Open General Intelligence: MiniMax H3 Is Now Open Source": https://www.minimax.io/news/minimax-h3-open-source. Accessed August 19, 2026.
- Wan-Video, official Wan 2.2 repository: https://github.com/Wan-Video/Wan2.2. Accessed August 19, 2026.
- Wan-AI, official Wan2.2-T2V-A14B model card: https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B. Accessed August 19, 2026.
- Alibaba Cloud Model Studio pricing: https://help.aliyun.com/en/model-studio/model-pricing. Accessed August 19, 2026.
- minimaxh3.tv model, generator, and pricing pages: https://minimaxh3.tv/models/minimax-h3, https://minimaxh3.tv/ai-video-generator/text-to-video, and https://minimaxh3.tv/pricing. Accessed August 19, 2026.
- Alibaba Cloud official Wan 2.2 video: https://www.youtube.com/watch?v=ktDogWm7Hac. Metrics captured August 19, 2026.
Last verified: August 19, 2026. This comparison uses current public documentation and interfaces. No new paid generation was performed, so it does not rank visual quality, success rate, latency, or prompt adherence.
