mırakı DEENFRIT
Miraki Mind

MiniMax H3 locally up to 4K: the refiner is open, the license has limits

NVIDIA’s Sana team just dropped the missing second stage of MiniMax H3 as an open refiner, and ComfyUI followed a day later. The finish can now run on your own hardware. Why you should still read the license first.

Miraki MindAI-assisted editorial · 2 November 2026 · 7 min read

On 27 October 2026, NVIDIA’s Sana team released “Open-H3-Regenerate-2K”, an open refiner for the video model MiniMax H3. It takes a low-resolution draft and redraws it in three compute steps, up to 2K or 4K. On 28 October, ComfyUI merged two matching building blocks into its main branch: a node for the H3 latent upscale and the fast decoder Sol-H3-VAE.

Where we left off with Pika H3 Flex

At the end of October we wrote about Pika H3 Flex here, an open model on the same H3 base. Our verdict then: strong for look tests and previz, but 768p is too low for a commercial, and the step up to 4K only ran on Pika’s servers. This is exactly the gap the new refiner fills.

There’s also a question we settle with every brand: what runs locally, what runs in the cloud? Scripts, faces and product data stay with us unless agreed otherwise. For final shots, though, we have so far still needed cloud models. The shots of It Ain’t Coming Off, for example, were made with Seedance 2.5.

A final shot from It Ain’t Coming Off, rendered in the cloud with Seedance 2.5: for this stage, H3 has had no open route so far

What’s actually new

H3 works in two stages. The base model fixes content and motion at low resolution, and the second stage redraws the draft in 2K. Only the first was open so far. The second, H3-Regenerate-2K, is available from MiniMax only as an API, without weights.

According to the project page, the open rebuild can do more than the original:

  • Input: drafts from 384p, including from other models such as LTX-2 or SANA-Video. The official API only accepts unmodified H3 clips in 768p.
  • Output: native 768p, 1080p, 2K (2560 × 1440) or 4K (3840 × 2160), for clips of 5 to 15 seconds in a single pass.
  • Speed: a 5-second clip from 768p to 2K takes 22.45 seconds on a B200 with NVIDIA’s Sol engine and fast decoder, instead of 87.26. The team doesn’t count the draft itself, loading the model or writing the file.

In ComfyUI, the way up is now core: the new node upscales the H3 latent directly, without a detour through decode and re-encode, and the audio passes through unchanged. Sol-H3-VAE decodes a 768p clip on an RTX 5090 in 1.7 instead of 10.7 seconds. Kijai, who submitted both changes, says it openly: a different, distilled decoder, not a lossless speedup, with differences showing in fine textures.

The fine print matters: according to kombitz’s weekly roundup, there are still no official ComfyUI nodes for the refiner itself, only conversions from the community. Hence “almost only” standard nodes.

The draft: content and motion are set, the details not yet. The finish: the refiner redraws the details at full resolution. The limit: the license excludes the EU, the United Kingdom, South Korea and the US.

What this means for a production

Draft first, then refine: that fits how we work, where the look gets approved first and the finish comes after. Locally, you could generate lots of low-resolution variants, and only the approved ones go to 2K or 4K.

The hardware remains a beast. According to Pika, the H3 base takes up around 210 GB. The refiner’s speed figures come from a B200, a data-center GPU. In fast mode, the decoder needs around 48 GiB of VRAM for 5 seconds in 2K; to make it fit on a 32 GB card, NVIDIA’s own ComfyUI variant moves the weights into system memory between requests. How long the whole route takes on a workstation, none of the sources say.

We do have our own experience with the model itself: we tested MiniMax H3 locally, on an RTX 5080. The result was so-so. We liked LTX 2.5 better.

What this means for a brand

Running locally means references and drafts stay on your own hardware, now including the finish. That’s exactly the part that still had to leave the building with the Pika model.

Honestly, the catch is the license. Refiner, fast decoder and base model run under the “MiniMax H3 Community License”. According to its text it applies worldwide, except in the EU, the United Kingdom, South Korea and the US, and it doesn’t allow using or showing the models or their outputs outside that territory. Switzerland isn’t on the exclusion list. For a campaign that also runs in Germany, France or Italy, it’s still an issue. Anyone who wants to work with it in the excluded countries should request a separate license from MiniMax.

On top of that come third-party components with their own licenses, such as the latent upscaler.

This bundle is not covered by a single blanket license.

Model card of Open-H3-Regenerate-2K

What this means for a specific job is for a lawyer to decide, not this post. Labelling doesn’t change: AI-generated material gets labelled, whether it’s rendered locally or in the cloud.

Our assessment and the limits

Technically, this is a real step: with H3, there is now an open route from draft to 4K finish, and two key building blocks sit in the ComfyUI core. The quality claims (better than the official API on RefinerBench-150, preferred in 87 percent of 200 blind judgments) are the team’s own numbers.

For brand work in Europe, the license holds things back more than the technology. Still open: how well the refiner holds faces and products from approved sheets across a whole clip, how long it takes on affordable hardware, and whether MiniMax changes the exclusion list. According to the license, MiniMax keeps reviewing the situation in these countries.

Sources

Have an idea that was too big?

Tell us about it in a few lines. We reply within one working day.hello@miraki.ch+41 31 348 90 90
Start a project