Downloading the H3 weights does not give you the H3 you have been watching

The MiniMax H3 weights are on Hugging Face and the download is not gated. That is real, and it is unusual for a video model at this quality level. It is also routinely reported as "H3 is open source", which sets up a specific disappointment: you pull 42 GB into ComfyUI, run the official template, and the result does not look like the clips that made you want the model.

Nothing is wrong with your setup. You have one third of a system.

The shape of the thing

H3 as a service runs in three stages:

your input (text / images / video / audio)
        ↓
① H3-Context-IR      hosted, not open
        ↓
② H3-Base            open weights  ← this is what you downloaded
        ↓
③ H3-Regenerate-2K   hosted, not open
        ↓
   the 2K file

Stage ② is the generator. It is genuinely open, it genuinely produces video with synchronised stereo audio, and it is the hard part of the problem. It is also the only stage you get.

① Context-IR is the part that reads your prompt properly

Context-IR takes whatever you gave it — a sentence, nine reference images, three video clips, an audio sample — and rewrites it into a structured intermediate representation the generator consumes. It resolves instructions across modalities, works out temporal ordering, and fills in semantics you left out.

The output is not a tidied-up version of your sentence. It is a shot-level document with fields:

• integrated_multimodal_description — per-shot visual description with timestamps like At 00:04.500
• overall_soundscape — ambience and foley
• non_diegetic_music — score, if any

Reference jobs get three more fields on top, including a per-asset retention analysis that says how strictly each reference must be held.

The token counts in the official examples show how much writing this stage does on your behalf: about 8,600 tokens for a text-to-video job, roughly 22,800 for image-to-video, and around 39,300 for a reference job. That is the document your one-line prompt becomes before the generator sees anything.

Run the open weights and none of it happens. Your sentence goes in as your sentence. The model is the same; the brief is a fraction of the size.

MiniMax's own guidance is explicit about the two ways out: call the Context-IR API, or build your own context layer following the published prompting guidance. There is no third option where the sentence is enough.

③ Regenerate-2K is not an upscaler

This is the part most worth understanding, because the name suggests something it is not.

A conventional upscaler receives a finished 768p frame and infers what the missing pixels should be. It has nothing else to work from, so fine detail — small text, distant faces, texture — is invented.

Regenerate-2K feeds the 768p result plus the original multimodal context back through H3 and generates again at 2K. It still has your prompt and your reference material, so detail that was too small to resolve at 768p can be reconstructed from what it was supposed to be rather than guessed from what it became.

That is why it is priced as generation ($0.05 per output second, with the source assets billed again) rather than as post-processing, and it is why "I will just upscale locally" is not the same trade.

What this means depending on what you are doing

There is a wrinkle in row two that is easy to miss. Since ComfyUI 0.33.1 there are partner nodes for Context-IR and Regenerate-2K, so a graph on your machine can reach 2K. Those nodes run on MiniMax's servers. A local pipeline that ends in 2K sends your 768p result and your source material off the machine, which undercuts the main reason most people self-host in the first place.

"Open weights", not "open source"

The distinction is not pedantry here, and the licence is where it bites.

The MiniMax H3 Community License restricts use to an "Applicable Territory" that excludes the European Union, the United Kingdom, the Republic of Korea and the United States. The restriction is written into the first sentence of the agreement, not into an appendix — and it extends to outputs, so a clip rendered somewhere permitted and published somewhere excluded is also outside the grant.

Two things are widely misreported about this:

1. The agreement is scoped to the Hugging Face repository. It governs the checkpoint. Someone who has never downloaded the weights and only ever called the hosted API has not entered into it — their relationship is with the service terms instead. The hosted API is documented as globally available.
2. Excluded regions have a formal route. MiniMax invites licence applications from them and frames the restriction as regulatory caution rather than policy: "The current limitation means 'not yet', not 'not ever.'"

Measured against the Open Source Definition, the licence fails two criteria — discrimination against persons or groups, and against fields of endeavour. The weights are open. The licence is not an open-source licence. Hugging Face's own label for it is other, so the publisher is not claiming otherwise either.

This is a summary, not legal advice. Read the LICENSE and the official Q&A before making a decision that depends on it.

Before you download 498 GB

One last practical note, since it catches people at the first step. The full repository is about 498.5 GB, because the raw checkpoints and the diffusers conversions sit side by side. You almost certainly want a subset:

Scope the download to what your framework needs. And if you are following a ComfyUI guide, check the version it targets — a great many of them still say 0.30.0, which shipped the nodes but not the nine H3 fixes that followed. Use 0.32.0 or later; two of those fixes address audio corruption and peak memory.