MiniMax H3 open weights: 768p locally, 2K only hosted
MiniMax's open H3 Base checkpoints top out at 768p. The 2K tier needs a proprietary module. What that means if you want to self-host or call H3 through Sume.

Can you run MiniMax H3 at 2K from the open weights? No. MiniMax's open-source announcement, dated 2026-08-03, says the base model outputs 768p by default and that 2K comes from a separate H3-Regenerate-2K module that is not yet open-sourced. The weights are for self-hosting; high resolution stays on the hosted side.
What is and is not released
The post releases the "MiniMax-H3 Base FL2VA" and "Base Ref2VA" checkpoints under the MiniMax H3 Community License, hosted on Hugging Face and ModelScope. It says H3-Context-IR is not included, and that content moderation applies.
| Part | Status per MiniMax |
|---|---|
| Base FL2VA checkpoint | Released, community license |
| Base Ref2VA checkpoint | Released, community license |
| H3-Regenerate-2K module | Not yet open-sourced; used for 2K |
| H3-Context-IR | Not included (proprietary) |
The hosted tier
MiniMax's hosted video generation guide lists H3 at 768P and 2K, 4 to 15 seconds. On Sume, minimax-h3 lists 480p and 768p as native tiers, with 2K and 4K priced as upscales at $0.13 and $0.16 per second list. Sume bills the list price times 1.25, rounded up to cents.
Sume's row starts at 5 seconds and takes up to 9 images, 3 videos and 3 audio clips, 12 files combined.
Decision
Self-host if you can accept 768p, have the GPUs, and need control over the weights under the community license terms; read that license first. Use the hosted route if you need 2K or want billing per second without running infrastructure. This post does not cover the license terms; read them first.
Questions to settle before self-hosting
If any answer is no, the hosted route is simpler. Upscaling a 768p local result with a separate tool is possible but is a different pipeline from the proprietary 2K module, and quality will differ.
- Does the community license permit your use, including commercial use? MiniMax's post names the license; this post does not summarise its terms.
- Is 768p enough for your delivery? Anything larger depends on a module that is not yet released.
- Can you absorb moderation behavior? The post says content moderation applies.
- Do you need the missing H3-Context-IR component? It is not included.
The announcement draws a line between what you can download and what MiniMax keeps running itself. That line is common in open releases, and it is the reason a headline like open source can mislead a buyer who needs a specific output size. Read the parts list first, then match each part to the resolution you need, and only then plan hardware.
On Sume, the choice is simpler: pick minimax-h3 for native 768p or minimax-h3-max for its listed tiers, and the provider runs the hosted pipeline for you.
Sources
Related posts
More in Models
- MiniMax H3 reference limits: 9 images, 3 videos, 3 audios on Sume
MiniMax H3 reference-to-video on Sume takes up to 9 images, 3 videos and 3 audio clips, 12 in total. The duration rules, the extra-image fee and a request.
- Nano Banana reference limits: Gemini slots vs Sume input_references
Google counts Nano Banana references by tier and by object versus character. Sume exposes one input_references range per row. How to read it and what to send.
- Native 4K AI video API: LTX-2.5 claims 4K, what Sume offers instead
LTX-2.5 is listed up to 4K but is not in Sume's catalog. Sume's 4K options are Gemini Omni Flash 1.1 and MiniMax H3 4K upscales. Limits and per-second prices.
- New AI video models since February 2026: dates and which Sume lists
Kling 3.0, Seedance 2.5, MiniMax H3 and LTX-2.5 launched in 2026. Release dates, length limits and which of them Sume's video catalog lists today.
Written by Sume