Chinese AI labs keep shipping video models faster than the leaderboards can settle — and this one shipped with a license that tells the US, EU, UK, and South Korea they can't run it locally.
Update — August 5, 2026: MiniMax's July 30 teaser is now a confirmed release. MiniMax published open weights for H3 on Hugging Face on August 3, 2026 — a 33-billion-parameter omni-modal model that generates 4-15 second clips at up to 2K/24fps with native stereo audio. The benchmark claims from the teaser held up: H3 ranks #1 in Video Editing (With Audio) on Artificial Analysis and places top-three in text-to-video and image-to-video. But the real story in the confirmed release is the license: MiniMax's Community License Agreement explicitly excludes the United States, the European Union, the United Kingdom, and South Korea from local deployment rights.
Update — August 11, 2026: The official repository now documents original BF16 checkpoints alongside Diffusers layouts and reference paths for SGLang, vLLM, Diffusers, and ComfyUI. It also added a portable h3-prompt-writing agent skill. Two August 9 patches supplied missing dependency pins and corrected audio-VAE weight initialization for models constructed from scratch; MiniMax notes that checkpoint inference was unaffected.
TL;DR
| Question | Answer |
|---|---|
| What is it? | 33B-parameter open-weight omni-modal video model — text, image, video, audio in; video with audio out |
| Released | Teased July 30, 2026; open weights confirmed August 3, 2026 |
| Generation specs | 4-15s clips, native 768p (2K via hosted API only), 24fps, native stereo audio |
| Local hardware | Official BF16 SGLang recipe: 4 GPUs; community quantized/offloaded paths can run on much less VRAM |
| How good? | #1 Video Editing (With Audio), top 3 in text/image-to-video per Artificial Analysis; ~1476 Arena.ai image-to-video score |
| License catch | Excludes the US, EU, UK, and South Korea from local deployment |
| Commercial ceiling | Requires separate written authorization above $20M/year revenue |
What H3 actually generates
H3 treats text, images, video, and audio as one unified input context rather than bolting audio generation onto a silent video model as an afterthought. From that combined input it produces clips with native stereo audio synced to the visuals — not a separately generated audio track stitched on after the fact.
The model supports meaningful reference-conditioning: up to 9 reference images, 3 reference video clips, and 3 reference audio clips per generation, enabling consistent characters and styles across multiple generated shots — a persistent gap in earlier open video models.
Two task-specific checkpoints ship separately: fl2va handles text-to-video and image-to-video, while ref2va is built specifically to maintain subject consistency across reference-conditioned generations.
Why "Video Editing (With Audio)" is the interesting ranking
Most AI video leaderboards split into text-to-video, image-to-video, and increasingly video editing — taking an existing clip and modifying it rather than generating from scratch. Audio-aware editing (matching dialogue, sound effects, or music to the edited output) is a much narrower category, one most competing models still treat as a separate pipeline bolted onto a silent video model.
H3's #1 ranking here — now independently confirmed rather than a teaser claim — puts it ahead of models that generate strong audio-video pairs from scratch but haven't focused specifically on editing an existing clip while keeping audio coherent. The #2-#3 rankings in Text-to-Video and Image-to-Video are more standard competitive positioning against a field that includes ByteDance's Seedance 2.5, Sora, Kling, and Runway.
Running it locally: what it actually takes
The official release now exposes original BF16 checkpoints and Diffusers-formatted components side by side. Community quantized builds remain the route for smaller machines:
| Configuration | Size |
|---|---|
| Community pruned + NVFP4 setup | ~42.5 GB |
| Community quantized variants for both tasks | ~63.4 GB |
| Original BF16 checkpoints | ~123.6 GB |
MiniMax's official SGLang launch command uses four GPUs with a Ulysses sharding degree of four. ComfyUI's community dynamic-offloading path puts the practical floor around 12GB VRAM with 64GB system RAM recommended, but that smaller setup is a quantized/offloaded compromise rather than the official BF16 reference. One resolution caveat remains: the local weights generate natively at 768px on the short edge; true 2K output requires the separately hosted H3-Regenerate-2K pass.
The other important boundary is system completeness. MiniMax released H3-Base, but the hosted H3-Context-IR preprocessing system, sparse-attention inference, and local H3-Regenerate-2K module are still missing. The repository's "full 2K" workflow therefore combines local H3-Base inference with MiniMax API calls; it is not a fully offline 2K pipeline.
The license: the actual story
This is the detail that matters more than any benchmark number. MiniMax's H3 Community License Agreement explicitly defines "Applicable Territory" to exclude the European Union, the United Kingdom, the Republic of Korea, and the United States of America. Running the open weights locally — using, modifying, or distributing outputs — isn't licensed in those four regions. The hosted API remains globally available; it's specifically the local, open-weight deployment path that's geo-restricted.
Additional terms worth flagging:
- Attribution required — any product built on H3 must prominently display "MiniMax H3" in its UI.
- Commercial ceiling — companies above $20M in yearly revenue need separate written authorization.
- No distillation — training a smaller model on H3's outputs isn't permitted.
- Separate licensing exists — MiniMax offers alternative commercial licensing by application.
This pattern — a strong open-weight release paired with a license excluding the largest Western markets from local deployment — is becoming a recognizable move among Chinese AI labs navigating export-control and IP dynamics. "Open weights" here is closer to "regionally available open weights" than a fully unrestricted release.
What people are asking
Is there now an official prompt-writing helper? Yes. MiniMax added one portable agent skill plus eight MiniMax Hub-specific video skills. The portable helper can be installed from the official repository:
npx skills add https://github.com/MiniMax-AI/MiniMax-H3 --skill h3-prompt-writing
The portable skill contains Markdown guidance and reference files rather than making its own API calls, so it can be read by Codex, Claude Code, Cursor, Windsurf, or another harness that understands SKILL.md files. The other eight skills depend on MiniMax Hub canvas tools and are not drop-in generic-agent packages.
Did the August 9 audio-VAE patch change generated videos? Not for normal checkpoint inference. The official commit says the broken initialization path only affected constructing the audio VAE from scratch; loading released weights through load_state_dict was unchanged. The same update added a top-level requirements.txt, closing a genuine day-one setup gap.
Can I still use the hosted MiniMax API from the US or EU? Yes — the territorial exclusion applies to running the open weights locally, not to MiniMax's own hosted API, which remains globally accessible under its normal terms.
How does this compare to Runway or Google's video models? Runway Aleph 2 and Google's Gemini Omni video stack are both closed, hosted-only products with no local-weights option — so for a US or EU team, the real comparison isn't "open vs closed," it's "closed-only" vs "open-but-geofenced." MiniMax's hosted API is the only currently available path either way in excluded territories.
Why exclude Korea specifically, alongside the US, EU, and UK? MiniMax hasn't published a stated rationale. The practical read is that these four jurisdictions have AI regulatory or IP-liability regimes MiniMax would rather not expose an open local-weights release to, pending separate licensing negotiations.
What was claimed vs. what's confirmed
The July 30 teaser and the August 3 release track closely, which is worth noting given how often pre-announcement claims from Chinese labs get revised on formal release:
| Claim (July 30 teaser) | Status (August 3 release) |
|---|---|
| "MiniMax-H3 is Open" | Confirmed — weights live on Hugging Face |
| #1 Video Editing (With Audio) | Confirmed by Artificial Analysis |
| #2 Text-to-Video, #2 Image-to-Video | Holds directionally — top-three range, some reordering |
| No license terms disclosed | Now disclosed: Community License, four excluded territories |
| No parameter count or specs | Now disclosed: 33B parameters, ~34.5GB-123.6GB depending on precision |
The gap that mattered most in the teaser — "open" with no license attached — is exactly the gap that turned out to have the biggest catch once the actual terms landed. That's a useful pattern to watch for with future teaser-to-release cycles from this and other labs: a benchmark claim tends to hold up reasonably well, but licensing details are the thing worth waiting for before making any adoption decision.
Where This Fits in the 2026 Video-Gen Field
MiniMax's Hailuo brand has spent 2026 competing directly against ByteDance's Seedance, OpenAI's Sora, and Kling for both consumer virality and enterprise API traffic — explainx.ai's complete guide to Sora, Runway, and Kling covers where each stood earlier this year. An open model with a category-leading edit-with-audio ranking is a meaningful differentiator in a field where most serious competitors (Sora, Kling, Runway) are closed APIs — provided you're outside the license's excluded territories.
What to verify before building on it
- Check the license text directly — MiniMax's H3 Community License Agreement, not a summary, before deciding whether your deployment region and revenue tier are covered.
- Choose the deployment class first — the official BF16 SGLang example uses four GPUs; the much smaller 12GB-VRAM claim refers to community quantization plus offloading. Verify both model size and runtime memory on your chosen framework.
- Don't assume 2K works locally — the open weights are native 768p; full 2K requires the hosted-only H3-Regenerate-2K pass, which is a separate API dependency even if you're running everything else locally.
- Re-check the leaderboard directly at artificialanalysis.ai/video/leaderboard/text-to-video — rankings shift as new models submit runs.
The takeaway
MiniMax H3 is a legitimately strong video model — the native-audio, reference-conditioning, and benchmark results are real, and "runs locally" is a genuine capability jump for open video generation. But the license is the actual news for anyone in the US, EU, UK, or South Korea: this is not the unrestricted open-weight release the "open source" framing implies for those markets, and any evaluation of H3 for production use needs to start with MiniMax's separate commercial licensing path, not the default Community License.
Related Reading
- antirez ported H3 to Apple Silicon in C and Metal — h3.c runs H3 locally on a Mac, and raises the territory-clause question
- FastGen-PDD: NVIDIA's 4-8 Step Distillation for Video and Image Models
- Seedance 2.5: ByteDance's 30-Second 4K AI Video Model
- Higgsfield's 33-Day Unlimited Seedance 2.5 Offer, Explained — H3 vs. Seedance 2.5, head-to-head
- AI Video Generation in 2026: Complete Guide to Sora, Runway, Kling
- ViMax: Agentic Video Generation Complete Guide
- Runway Aleph 2 vs Gemini Omni Video
- Liquid AI LFM2.5-2.6B: On-Device Agent Model
- Top Chinese AI Companies and Startups Guide
Sources: MiniMax H3 on Hugging Face, official GitHub repository and update history, MiniMax official announcement on X, Artificial Analysis video leaderboard. Specs, integrations, and license terms were rechecked August 11, 2026 — verify current license terms directly before any commercial deployment.