Is MiniMax H3 Max open source? What the H3 card says and Sume lists
MiniMax H3 has a Hugging Face card under a community licence. The card lists no H3 Max weights. Compare the two Sume rows with one script.

Not as far as the vendor's page shows. The MiniMax H3 card on Hugging Face (read 2026-10-03) describes released H3 checkpoints under a community licence, and it describes the H3-Regenerate-2K model as not yet open-sourced. It does not mention an H3 Max checkpoint.
Sume lists both minimax-h3 and minimax-h3-max as hosted rows. That tells you what you can call, not what is downloadable.
What does the H3 card release?
From the card as read on 2026-10-03:
- A 33B-parameter model with FL2VA and Ref2VA checkpoints.
- A community licence agreement, with an application form for users in the USA, EU, UK and South Korea.
- An SGLang example that uses four GPUs.
- H3-Regenerate-2K marked as not yet open-sourced.
Why is Max not on that list?
Max is a separate model name. Sume's video docs call minimax-h3-max the faster 768p variant, with 1080p coming from a latent refinement step. A faster variant under a different name is not covered by a card that never names it, so treat its weights as unconfirmed until MiniMax publishes them. Do not infer them from the base card.
How do you compare the two Sume rows?
The catalog shows what Sume exposes for each id. This script prints durations, resolutions and the hugging_face_id field for both:
import os
import requests
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
rows = requests.get("https://api.sume.com/v1/videos/models", headers=H, timeout=30).json()["data"]
for m in rows:
if m["id"] in ("minimax-h3", "minimax-h3-max"):
print(m["id"])
print(" durations:", m.get("supported_durations"))
print(" resolutions:", m.get("supported_resolutions"))
print(" hugging_face_id:", m.get("hugging_face_id"))What should you do if you need weights?
If you need to self-host, you can apply for the H3 release under its licence, and you can run the hosted rows meanwhile to judge the output. If you need the Max behaviour specifically, hosted is the only route the sources support today.
| Model | Weights on the H3 card | Sume row |
|---|---|---|
| H3 (FL2VA, Ref2VA) | Released under the community licence | minimax-h3 |
| H3-Regenerate-2K | Described as not yet open-sourced | Not a separate Sume id |
| H3 Max | Not named on the card | minimax-h3-max |
Sources
Related posts
More in Models
- Kyutai Pocket TTS languages: what it speaks vs hosted Sume TTS
Pocket TTS is a 100M-parameter open model you run yourself. Its README lists seven languages; Sume's TTS is hosted, per character, with a voice id.
- Pocket TTS languages: six or seven, and Sume's language field
Kyutai lists six Pocket TTS languages on its model card and blog, seven in the GitHub README. Here is how to read that, and how Sume TTS sets a language.
- Omni Flash vs H3 Max for reference-to-video: limits side by side
Gemini Omni Flash 1.1 takes 10 images and 3 short videos; MiniMax H3 Max adds audio refs, 12 files total. Limits, tags and a sample request on Sume.
- Shortest AI video clip via API: 2 seconds on Wan 3.0 only
Sume's video catalog by minimum duration: Wan 3.0 takes 2 seconds, Omni 3, Seedance and Kling 4, MiniMax H3 5. Plus how to trim below a floor.
Written by Sume