Which image and video models on Sume have open weights?
Sume's catalog is hosted. Where vendors publish weights for families Sume lists (Qwen-Image, MiniMax H3, FLUX.2) and why a null hugging_face_id proves nothing.

Sume does not publish a per-model open-weights flag, so the answer comes from the vendors. Three families on Sume's catalog have weights somewhere on a vendor page: Qwen Image (Apache 2.0 weights for the original Qwen-Image), MiniMax H3 (a community-licensed release, with the Max variant not on the card), and FLUX.2 (only the dev model is open, and it is not the one Sume lists). Wan 3.0 and Ideogram V3 have no weights on the vendor pages I read.
None of that says what Sume runs behind a model id. Sume's docs say upstream provider identity stays internal, so a family having open weights does not mean your call uses them. All vendor facts here were read on 2026-10-03.
Which Sume rows have vendor-published weights?
The table matches Sume's catalog ids from the Image and Video docs against what each vendor's own page shows. "Not found" means not on the page I read, not proof that none exist.
| Sume id | Vendor weights on the vendor page | Licence as shown | Caveat |
|---|---|---|---|
| qwen/qwen-image | Qwen/Qwen-Image on Hugging Face, 20B parameters, released Aug 4 2025 | Apache 2.0 | Sume does not say which build serves the id |
| black-forest-labs/flux.2-pro and flux.2-flex | BFL lists pro and flex as API only; FLUX.2 dev is the open-weights model | Dev is non-commercial per BFL's pricing page | Dev is not a Sume id |
| ideogram/ideogram-v3 | Ideogram's Hugging Face organisation lists only Ideogram 4 repositories | Ideogram 4 is non-commercial | V3 weights not found; Ideogram 4 is not a Sume id |
| minimax-h3 | MiniMaxAI/MiniMax-H3: 33B parameters, FL2VA and Ref2VA checkpoints | MiniMax H3 Community License Agreement | Sume does not say it runs the released checkpoints |
| minimax-h3-max | Card lists H3-Base; H3-Regenerate-2K is described as not yet open-sourced | Same community licence | No Max weights found on the card |
| wan-3.0 | Wan-AI organisation lists Wan2.2 models; no Wan 3.0 or 2.7 weights | No licence shown on the organisation page | The open family is older than the hosted one |
Why is hugging_face_id null on every Sume video model?
The Video docs show each row of GET /v1/videos/models with a hugging_face_id field, and the sample value is null. That field mirrors an OpenRouter field; Sume's catalog leaves it empty. Reading null as "closed model" is wrong: MiniMax H3 has a public card and the field is still empty. Use the vendor's page to answer the weights question, and use the catalog to answer what you can call.
Does the licence on the weights apply to a Sume call?
Generally no, because a hosted call is governed by the terms of the service you use, not by the licence on a download you never made. The weights licence matters when you run the files yourself. The practical exception is a model you chose because of its licence: if you want Apache 2.0 weights so you can fine-tune, a hosted id does not give you that, since Sume's Image API has no field for custom weights.
- Want to fine-tune: download the weights and host them; Sume cannot take them.
- Want to ship client work without a licence negotiation: use a hosted id and read the service terms.
- Want to test before downloading 20B or 33B parameters: call the Sume row for the same family first, accepting that the build may differ.
How do you test an open-weights family before you download it?
Use the hosted row as a probe. It will not match your local build exactly, but it shows whether the family handles your prompts. This script lists the catalog rows that map to the families above, with the fields that matter for a test:
import os
import requests
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
img = requests.get("https://api.sume.com/v1/images/models", headers=H, timeout=30).json()["data"]
vid = requests.get("https://api.sume.com/v1/videos/models", headers=H, timeout=30).json()["data"]
for m in img:
if any(k in m["id"] for k in ("qwen", "flux", "ideogram")):
print("image", m["id"])
for m in vid:
if any(k in m["id"] for k in ("wan", "minimax")):
print("video", m["id"], m.get("supported_durations"), m.get("hugging_face_id"))Sources
- Hugging Face: Qwen/Qwen-Image (read 2026-10-03)
- Hugging Face: MiniMaxAI/MiniMax-H3 (read 2026-10-03)
- Hugging Face: Wan-AI organisation page (read 2026-10-03)
- Hugging Face: ideogram-ai organisation page (read 2026-10-03)
- Black Forest Labs: FLUX.2 models (read 2026-10-03)
- Video generation
- Image API
- Image API reference
Related posts
- Hugging Face id on Sume video models: the field is always null
- Qwen Image license: is Qwen-Image open source under Apache 2.0?
- MiniMax H3 open weights on Hugging Face: what is in the release
- Wan 3.0 open source: weights on Hugging Face or API only?
- Open-weights video or hosted API? Four checks, Wan, LTX, Runway
More in Comparisons
- Wix Stores promo video maker vs a custom AI product clip
Wix Stores can auto-make a promo video per product, free. When is a custom Sume image-to-video clip from your own photo worth the extra step? A side-by-side.
- YouTube reused vs inauthentic content: which rule hits AI Shorts?
Two YouTube monetization rules and one new Shorts reach change, side by side. Which one an AI-generated Short can trip, and what to vary to avoid each.
- Sume vs Argil: AI avatar video and video agents compared
Argil makes AI-avatar and story videos with a chat agent, Director; Sume is a video agent with a multi-model API. Avatars, API, pricing, and limits compared.
- Sume vs fal: a generative media API or a video agent platform
fal runs 1,000+ image, video, and audio models behind one API. Sume adds a video agent, Formats, and avatars to a multi-model API. How the two surfaces differ.
Written by Sume