MiniMax H3 open weights: what is released and what stays hosted

MiniMax released the H3 base model with two task checkpoints on August 3, 2026, but it kept two modules hosted: the Context-IR preprocessing step and the 2K regeneration step. If you want the 2K-class output that MiniMax advertises, you use the hosted API, not the open files.

Sume does not run the open checkpoints. It routes minimax-h3 and minimax-h3-max through its video catalog and bills list x 1.25, so the open-weights question is mostly about what you could build locally versus what you can order per clip.

The parts list, from MiniMax's own page

MiniMax's open-source news post (read 2026-10-05) lists what is open: the H3-Base 33B-parameter Omni-Transformer, an FL2VA checkpoint for text-to-video and first/last-frame, a Ref2VA checkpoint for reference-to-video, and the supporting processor, tokenizer, text encoder and visual and audio VAEs. The license is the MiniMax H3 Community License Agreement.

What the 768p default means

The same page says the open pipeline sets the shorter side to 768 pixels by default and that 2K is not yet open. The page lists a maximum output of 15 seconds at 24 fps with stereo audio. That lines up with the shape Sume exposes: 5 to 15 seconds and native 768p.

Where Sume sits relative to the open files

Sume's docs call minimax-h3 the model with native 480p/768p, priced upscales for 2K/4K, and minimax-h3-max the faster 768p variant with a 1080p latent refinement. Neither is documented as a self-hosted artifact. The decision for a team is simple arithmetic on one side and infrastructure on the other.

On the hosted side, a 10-second 768p clip is $0.75 on minimax-h3 ($0.06 list x 1.25 = $0.075 per second) and $1.00 on minimax-h3-max ($0.08 list x 1.25 = $0.10 per second). On the open-weights side you carry GPU time, queueing, the Ref2VA and FL2VA servers, and the license terms, and you do not get the hosted 2K module.

• Prototype locally if you need weights-level control and can live with 768p.
• Order from the catalog if you need 1080p refinement, 2K-class upscales or no GPU operations.
• Read the community license page before shipping outputs commercially; this post does not summarize its territory terms.

A short decision rule

Pick open weights when the constraint is control over the model files. Pick the Sume rows when the constraint is time to a finished 5 to 15 second clip with audio. Before you integrate, call GET /v1/video-router/models and check the row for the resolutions it actually lists today.

What this changes for planning

If you were hoping that open weights would give you a free 2K pipeline, the vendor page says otherwise: the regeneration module that produces the larger frames is hosted only. A local pipeline at 768p can still be useful for previews, for fine-tuning experiments, and for teams that must keep inputs on their own machines.

If you were hoping to compare a local run with the Sume row, compare like with like. A 10-second clip at 768p, 24 fps, with stereo audio is the common denominator. On Sume that is $0.75 on minimax-h3 and $1.00 on minimax-h3-max for the clip, with no extra for audio because the audio is always on.

What Sume does not claim

Sume's docs do not describe H3-Base, the Ref2VA or FL2VA checkpoints as products. They describe catalog ids with durations, resolutions and prices. This post uses the MiniMax pages only for the release facts and the Sume docs only for what the router accepts. The license is MiniMax's, and territorial or revenue conditions in it are not restated here.