MiniMax H3 open weights: what is released and what stays API-only
MiniMax opened H3-Base on 2026-08-03 with two checkpoints, but 2K regeneration and Context-IR stay API-only. How that maps to Sume's minimax-h3 rows.

MiniMax released the H3-Base weights on August 3, 2026, with two task checkpoints, but two parts of the H3 system stay behind its API: the Context-IR preprocessing and the 2K regeneration step. So "H3 is open" means the core video-audio transformer is downloadable; 2K output and the instruction preprocessor are not.
If you want H3 without running a 33B model, Sume lists minimax-h3 and minimax-h3-max as pay-per-second rows. Everything on this page was read on 2026-10-07 from MiniMax's open-source announcement, its H3 post and Sume's video docs.
The release, piece by piece
MiniMax's launch post of July 31 describes the model as taking text, images, video and audio as one context, and returning video with native stereo sound at up to 15 seconds.
| Piece | Status | Detail from MiniMax |
|---|---|---|
| H3-Base transformer | Open weights | 33B-parameter dense, single-stream Transformer |
| FL2VA checkpoint | Open weights | Text and frame to video |
| Ref2VA checkpoint | Open weights | Reference to video |
| H3-Context-IR | API only | Multimodal instruction preprocessing |
| H3-Regenerate-2K | API only | High-resolution enhancement |
| Output | As stated | 4 to 15 s, 24 FPS, 768p locally and 2K via API, 32 kHz stereo audio |
| License | As stated | MiniMax H3 Community License Agreement |
What that means for a buyer
- Downloading the weights gets you 768p locally; the 2K step needs MiniMax's API.
- The license is a community license, so read it before commercial use.
- A self-hosted run needs your own GPU and the two checkpoints are separate tasks.
The Sume rows for H3
Sume's docs list minimax-h3 at native 480p and 768p for 5 to 15 seconds, with 2K and 4K as billed upscales when requested, and minimax-h3-max at 480p, 768p and 1080p (the 1080p is a latent refinement from native 768p). Both always produce native stereo audio. Sume bills provider list times 1.25.
| Row | Resolution | List per second | Sume per second |
|---|---|---|---|
| minimax-h3 | 480p | $0.05 | $0.0625 |
| minimax-h3 | 768p | $0.06 | $0.075 |
| minimax-h3 | 2K upscale | $0.13 | $0.1625 |
| minimax-h3 | 4K upscale | $0.16 | $0.20 |
| minimax-h3-max | 480p | $0.05 | $0.0625 |
| minimax-h3-max | 768p | $0.08 | $0.10 |
| minimax-h3-max | 1080p | $0.16 | $0.20 |
Pick by the missing piece
If you need the 2K step, the open weights do not include it, and on Sume the 2K path is the minimax-h3 upscale. If you want to keep prompts and references on your own machine, the weights are the route and you carry the GPU bill. If you want a draft at 768p for ten seconds, that is $0.75 on minimax-h3 and $1.00 on minimax-h3-max.
Open weights and a hosted row are different products
The decision between downloading weights and calling a row is mostly about who runs the machine. Weights give you control of data flow and a fixed hardware bill, and they give you the work of serving two checkpoints, keeping drivers current and testing every update. A hosted row gives you a per-second price, a catalog entry that Sume keeps current and queueing handled for you.
Neither route changes what the model can do at the pixel level. MiniMax's note says 768p is the local output and 2K comes through its API, so a pipeline that must deliver 2K cannot be satisfied by the weights alone.
- Choose weights for data control and a steady heavy workload.
- Choose a hosted row for bursts, drafts and no GPU to run.
- Read the community license before commercial use of the weights.
A rough way to think about volume
At $0.075 per second on Sume's minimax-h3 at 768p, a thousand 10-second clips cost $750. Whether that beats owning GPUs depends on your own hardware cost and utilization, which this page cannot know. Compute the break-even from your own numbers, and keep the hosted row for the weeks when volume is low.
Sources
Related posts
More in Models
- Where can you use Gemini 3.8 Flash TTS today? API, Notebook, Vids
Google lists Gemini 3.8 TTS in the Gemini API, AI Studio, Gemini Notebook and Google Vids now; the Enterprise API is coming soon. What Sume offers instead.
- Which Sume image models are text-to-image only, with no photo edits
Five Sume image models reject reference images: Higgsfield Soul, Imagen 4 Fast, Imagen 4 Ultra, Recraft V4 and Qwen Image Max. Prices, ratios and edit options.
- Which Sume video clips carry audio for the Shorts editor split
YouTube Shorts now separates audio and video tracks. Which Sume video models return sound, which can mute, and what a $0.01 detach adds.
- Which Sume video models take a source video_url? Edit, transfer
Only three Video Router rows accept a video_url source: gemini-omni-flash-1.1 (edit), higgsfield-genjutsu (motion transfer) and h3-max-recast (person swap).
Written by Sume