---
title: "Discussion on \"What MiniMax H3 actually outputs: \"1080P\" is 1920×1056, durations snap to 24 fps\" | Hashnode"
description: "Discussion on \"What MiniMax H3 actually outputs: \"1080P\" is 1920×1056, durations snap to 24 fps\". Ten measured H3 runs: pixel caps, 24 fps duration rounding, native audio, and what each tier really costs per second."
url: "https://hashnode.com/posts/minimax-h3-output-specs/6ae314cb2dce0cf071580454"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [What MiniMax H3 actually outputs: "1080P" is 1920×1056, durations snap to 24 fps](https://callirra.hashnode.dev/minimax-h3-output-specs?utm_source=hashnode&utm_medium=feed)

Ten measured H3 runs: pixel caps, 24 fps duration rounding, native audio, and what each tier really costs per second.

Oct 29 · 7 min read · callirra.hashnode.dev  
Read the article

-   [#minimax](https://hashnode.com/tag/minimax)
-   [#text-to-video](https://hashnode.com/tag/text-to-video)
-   [#image-to-video](https://hashnode.com/tag/image-to-video)
-   [#video-generation](https://hashnode.com/tag/video-generation)
-   [#ai-video](https://hashnode.com/tag/ai-video)

## Written by

[Callirra](https://hashnode.com/@callirra)

@callirra

  
[View profile](https://hashnode.com/@callirra)

One workspace for AI image, video and audio

Callirra is one workspace for AI image, video and audio: 50 media models, 11 multimodal chat models, prompt polish, a private asset library, saved Constellation workflows and AI Filmmaker. The team writes here about real credit costs, model comparisons and release notes. Product and pricing: callirra.com

## Responses 3

Markdown supported · ⌘↵ to post

[Viktor Novak](https://hashnode.com/@viktornovak)

[Oct 29](https://hashnode.com/posts/minimax-h3-output-specs/6ae314cb2dce0cf071580454/comment/6ae31bdccecd765f9d38c65f)

The 32 kHz audio track is the detail I would have missed. We hit the same thing on a different lane - every clip came back with an ambient bed nobody asked for, and stripping it after the fact cost an extra encode pass. Now we check the stream list before the first edit. The two-pass 2K economics also match what we measured: the regeneration is billed twice, so the headline 2K rate is never the real number.

[Yuki Tanaka](https://hashnode.com/@yukitanaka)

[Oct 29](https://hashnode.com/posts/minimax-h3-output-specs/6ae314cb2dce0cf071580454/comment/6ae317258726f12d1091686b)

The 1056 thing is going to quietly break a lot of overlay templates. Anyone authoring at 1920x1080 gets a 24px vertical mismatch that only shows up after the first export.

[Camille Dubois](https://hashnode.com/@camille-dubois)

[Oct 29](https://hashnode.com/posts/minimax-h3-output-specs/6ae314cb2dce0cf071580454/comment/6ae31a862a881abf269d0e1c)

The 24 fps quantisation is the part I keep thinking about. 1.625s for a requested 1s is a 62% overshoot - that is not a rounding error, that is a different shot. For beat-synced work the only clean path I can see is generating longer than you need and cutting to the timeline, instead of asking the model for an exact duration. Curious whether the 39-frame result is stable across retries, or whether the frame count itself drifts run to run.
