Download README.md from alphabell/MiniMax-Video-Budget: direct link, hf CLI and curl.
- Browser
- Download file 2.55 kB
-
https://huggingface.co/spaces/alphabell/MiniMax-Video-Budget/resolve/main/README.md
- Command line
-
hf download hf://spaces/alphabell/MiniMax-Video-Budget/README.md
-
curl -L -o README.md https://huggingface.co/spaces/alphabell/MiniMax-Video-Budget/resolve/main/README.md
title: MiniMax Video Budget
emoji: 🎞️
colorFrom: green
colorTo: gray
sdk: static
app_file: index.html
license: mit
short_description: Video frame-to-placeholder budget for MiniMax M3.1 Flash
thumbnail: >-
https://huggingface.co/spaces/alphabell/MiniMax-Video-Budget/resolve/main/cover.png
models:
- MiniMaxAI/MiniMax-M3.1-Flash
tags:
- minimax
- video
- multimodal
- tokenization
- visualization
- developer-tools
MiniMax Video Budget
How many vision placeholders does a sampled video create in MiniMax M3.1 Flash?
Open the interactive inspector to enter the number of sampled frames and their post-resize dimensions. The page shows two-frame groups, odd-frame repetition, spatial patch merging, a shareable URL and a JSON export.
The calculation is ceil(sampled_frames / 2) × (processed_height / 28) × (processed_width / 28). For 16 sampled frames already processed at 672 × 672 pixels, that is 8 groups × 576 placeholders = 4,608 vision placeholders. A single frame also occupies one temporal group: its final frame is repeated, yielding 576 placeholders at this size.
This is a source-derived count of repeated vision placeholders. It is not total prompt tokens, memory, speed, video quality or a prediction of automatic frame sampling or resizing. Timestamp and wrapper tokens are additional. No video, image or model weights are uploaded or run by this Space.
The formula follows the pinned video processor, placeholder expansion, image merge setting and model card. The exact revision and checked source hashes are in sources.json.
Built by alphaBell as an independent educational tool. MiniMax publishes the model and processor. No publisher source code or model weights are included. Source attribution · Viewer license.
Serve this directory over HTTP to run locally. Run node --test tests/*.test.cjs for calculation and release checks. The viewer uses no backend, analytics, font CDN or API key.