MiniMax H3 Uncensored — The 2K Stereo-Audio Model and What It Does With Adult Prompts

What MiniMax documents

MiniMax’s launch article describes H3 as an open model that understands unified context across text, images, video and audio, generates video with native stereo sound, and outputs up to 15 seconds at 2K resolution, with 768p as a cheaper tier. The company claims 2K per-second pricing under a third of mainstream models and 768p under half of mainstream 720p. Reference prompting is conversational: “reference the camera movement from Video 1, have the character in Image 2 sing with vocals matching Audio 3.” That reference grammar is the model’s signature.

What Studio exposes

Resolution at 768P or 2K, duration from 4 to 15 seconds, and ratio from seven options (adaptive, 16:9, 4:3, 1:1, 3:4, 9:16, 21:9) for text-to-video, adaptive for image-to-video. Image-to-video takes one start frame. Reference-to-video takes up to 9 reference images, 3 video clips and 3 audio clips, 12 assets total. The default quote at review time was 1,110 credits; MiniMax bills input and output video seconds at the selected resolution rate and includes the first five reference images, so long reference clips raise the quote.

Early adult-content data, stated plainly

Sixty days to 9 September 2026: 7 MiniMax H3 reference-to-video jobs from 3 creators, 2 finished, 4 stopped by the content-safety response, 1 other failure. Seven jobs is far too few to publish a rate, so we will not; it is enough to say the upstream route does review adult content and has refused some of it. For comparison, Wan 3.0 reference-to-video finished 94% of 6,901 jobs and Wan 2.2 Spicy logged zero stops in 16,059. We will update this section once H3 has a few hundred runs behind it.

Where H3 is worth choosing today

Dialogue-led clips: H3’s stereo audio is generated in the same pass and its reference grammar makes “match the voice in Audio 1” a one-line instruction. Music or rhythm-synced motion. Suggestive, clothed character work at 2K where texture matters, and 21:9 cinematic framing. Its 768P tier is a cheap way to iterate an idea with sound before rendering the same brief on Wan 3.0 at a higher tier if the adult content is too explicit for H3’s review.

Deployment filtering and adult-content boundaries

Makeartly does not add a separate mainstream adult-content refusal layer for verified adult creators; this deployment-level statement does not mean the upstream official model is uncensored or guarantee prompt acceptance. Adult workflows are for adults only. CSAM and all under-18 sexual content are prohibited. Consent and likeness authorization are required, and non-consensual intimate content, unauthorized real-person deepfakes, and sexual impersonation are prohibited.

A pairing that works

Use H3 for the parts of a project that lean on audio and 2K texture and stay implied or clothed; use Wan 3.0 for explicit multi-reference scenes and Wan 2.2 or 2.7 Spicy for explicit single-still clips. Both H3 and Wan 3.0 run in the 1,100-credit range for a default clip, so the choice is about which review your brief will pass and whether you need 30 seconds (Wan) or stereo sound at 2K (H3).

A 12-second clip where she speaks two lines to camera in stereo

MiniMax H3 is the strongest tool on the platform for this; keep the reference clothed and quote the lines.

An explicit multi-reference scene

Use Wan 3.0 reference-to-video; H3’s upstream review has already refused adult attempts in its first weeks.

Acceptance test H3 at 768P with a clothed reference

Run one 5-second 768P clip as the acceptance test, compare the voice and 2K-tier texture against a Wan 3.0 render of the same brief, stop at the first review refusal, and choose Wan 3.0 for the explicit remainder of the project.

The H3 adult data on this page is seven jobs; treat every claim about its permissiveness, ours included, as provisional. Reference seconds are billed: MiniMax bills input video seconds as well as output, so a 15-second reference clip roughly doubles the quote of a 15-second render.