MiniMax H3 Free: Two Open Checkpoints, a Closed 2K Module, and Four Excluded Markets
MiniMax H3's weights are open and royalty-free inside a territory that excludes the EU, the UK, South Korea and the United States; the module that produces 2K output is not in the release at all, and the official API publishes per-second prices with no free seconds while its video resource packs explicitly exclude H3. Read 2–3 November 2026.
- Model
- MiniMax H3
- Observed
"Free" on MiniMax H3 is three separate offers, and only the first one is unconditional. The weights are open: two task checkpoints, 144.05 GB for a single task family, royalty-free inside a territory that excludes the European Union, the United Kingdom, South Korea and the United States. The module that turns a 768-pixel result into 2K is not in the release. And the official API, which is where the 2K path actually runs, publishes a per-second price and no free seconds at all.
The three are usually collapsed into one word by the videos that rank this model, which is why the numbers below are worth separating. Each one answers a different question: what can you run, what can you finish, and where are you allowed to use it.
Everything below comes from primary pages read on 2 and 3 November 2026, without a login and without spending a credit: the model card and the licence file on Hugging Face, the repository's file listing through the Hugging Face API, MiniMax's own pricing page, and the ComfyUI tutorial for the model. No generation was run for this page, on any hardware. That limit is stated again at the end, where it matters.
What is actually in the download
The repository hosts two task checkpoints and their diffusers layouts side by side. The card describes them precisely enough to plan around:
| Checkpoint | What goes in | What comes out |
|---|---|---|
H3-Base-FL2VA | text, plus zero, one or two images as first and last frame | video and audio |
H3-Base-Ref2VA | text, plus up to 9 images, up to 3 video clips and up to 3 audio clips | video and audio |
The reference mode carries the tighter ceilings, and they are the ones that decide whether a workflow fits: each video clip must be 2 to 15 seconds, the videos together may not exceed 15 seconds, the same limits apply to audio, and the total number of files across every input type is capped at 12. Output is 4 to 15 seconds, 24 frames per second, with the shorter side set to 768 pixels by default.
The size is the part that rarely makes it into a tutorial. Summing the file listing through the Hugging Face API gives 144.05 GB for one task family and 498.5 GB for the repository as published, across 280 files, because both families ship full BF16 weights and each carries its own video VAE. Those two figures are this site's arithmetic over a published file list, not numbers MiniMax prints anywhere.
The 2K module is the one that stayed closed
The card is unusually direct about what did not ship. Three pieces are named as missing, and they are not equivalent.
H3-Context-IR is the first: "Because H3-Context-IR relies on a multi-stage workflow and multiple hosted models and services, it is not included in this open-source release." The card does not treat that as a detail. It says the module is "critical to the quality of the final output" and recommends building your own context-processing system from the prompting guidance if you are not using the API.
H3-Regenerate-2K is the second, and this is the one that decides the resolution question: "Due to the complexity of the system, this module is not yet open-sourced. We will release it once it is ready. We provide an API for validating the official results." The base checkpoints produce 768 pixels on the short side; 2K is a second pass that feeds that result and the original context back through the model.
The third is smaller but worth recording: the sparse-attention implementation used in the final training stage is "not included in the initial open-source release and will be published separately in a future update."
So the honest description of the download is a 768-pixel engine with an open base and a closed finishing stage. The ComfyUI tutorial says the same thing in one sentence: local runs give "about 1 megapixel (768px short edge) at 24fps for up to ~15 seconds", and "2K is a hosted-service output and needs a separate upscale pass in ComfyUI."
The API has prices, not free seconds
MiniMax's pricing page lists the model in two resolution bands, and both are per second: ¥0.80 at 2K and ¥0.50 at 768P. Input images are priced separately at ¥0.20 each once the free image allowance is used up. There is no free-video grant on that page for H3, and nothing resembling a trial quota.
The more consequential line is in the resource-pack section, and it appears six times across the page in different renderings: "视频资源包支持 Hailuo 系列视频模型，暂不支持 MiniMax H3" — the video resource packs cover the Hailuo series and explicitly do not cover MiniMax H3. That closes a route that exists for MiniMax's other video models: prepaying for a discounted pack. On H3 the meter runs at list price.
For anyone comparing this to a subscription, the shape matters more than the number. A model with no pack and no free seconds is billed per second from the first second, and the 2K path — the one the open release cannot finish — is the more expensive of the two bands.
The licence decides where "free" applies
The community licence is dated 2 August 2026, and its scope is territorial before it is anything else: "The scope of this License Agreement is expressly limited to the 'Applicable Territory'."
The definitions do the work. "Applicable Territory" means worldwide excluding the Excluded Territories, and the Excluded Territories are "the European Union, the United Kingdom, the Republic of Korea and the United States of America." Inside the applicable territory the grant is generous — "non-exclusive, non-transferable, royalty-free, limited license to use, reproduce, distribute, create derivative works ... and modify the Materials" — and outside it the licence is not merely quiet, it is prohibitive: "You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory. Any such use outside the Applicable Territory is not authorized by this Agreement."
Two commercial conditions sit on top. A business whose products generate more than 20 million US dollars a year needs prior written authorization from MiniMax, requested by email with a fixed subject line. And any commercial product using the model must "prominently display 'MiniMax H3'" on its interface.
That is the layer the tutorials skip. A reader in Berlin, London, Seoul or anywhere in the United States can download the weights and run them; what the licence says about that use is a separate question, and it says the use is not authorized. The licence does leave a door open — it invites anyone in those territories to ask for a licence "granted based on robust controls and guardrails" — but that is a negotiation, not a download.
There is also a reseller in the picture. Comfy's licence page describes itself as "the only official reseller of MiniMax commercial-use licenses, covering MiniMax H3 plus Audio & Music". The page's tier prices were not readable in the copy saved here, so this page records the claim and not a number.
What people run instead, and what it costs them
The workaround the community has settled on is a rented GPU, and the numbers in the comments are hardware numbers rather than licence numbers. One commenter reports "around 11 minutes for a 10-second video at 1mp with the turbo lora. (16 GB RTX 4080 & 96 GB system RAM)". Another prices the alternative: "I rent an RTX Pro 6000 Blackwell with 96GB VRAM on Runpod for about $1.69 per hour instead of buying the card, which currently costs around $16,000 by itself."
Both are user reports from a single video's comment section, captured on 2 November 2026, and neither was reproduced here. They are also hard to square with the published size: 144.05 GB of BF16 weights does not fit in 16 GB of video memory, so the first report describes a quantised or heavily offloaded setup, and how it was assembled is not explained in the comment.
What this does not settle
No clip was generated for this page, on a rented card or anywhere else. Whether the open checkpoints produce good video, whether the missing context module degrades quality in practice, and how far a quantised local run drifts from the hosted one are all unmeasured here. Reading a licence and a file list tells you what you may run and what it weighs; it does not tell you what comes out.
The 144.05 GB and 498.5 GB figures are this site's addition over the repository's file listing, and file listings change. They describe the repository on 2 November 2026, not a permanent property of the model.
The territorial reading is a reading of the licence text, not legal advice, and the licence itself says MiniMax will keep evaluating the excluded markets. Anyone whose revenue or location sits near one of those lines should read the agreement rather than this page.
Prices, packs and licence terms move without announcement. The readings behind this page are dated 2 and 3 November 2026 and carry the page each was taken from; re-read before quoting.
Related on this site: Wan 3.0 free — the same question asked of a model whose weights are not open, where the free allowance is a metered 30 seconds. Seedance 2.5 API free — how a token grant and a mandatory resource pack coexist on one pricing page. Hailuo AI free — the credit rate that H3's resource packs still serve.