Pair the frames
The processor pads an odd frame count by repeating its last frame, then groups frames in twos.
Source-pinned model lens 01 / MiniMax M3.1 Flash
One video.
Many visual slots.
MiniMax M3.1 Flash groups sampled frames in pairs, then expands each group into visual placeholders. Enter the frame count and processed frame dimensions to see the exact arithmetic.
Processor arithmetic, not an inference benchmark
01 / Try a video shape
Count the visual placeholders
Use the number of frames after sampling and the height and width after resizing. Dimensions must be multiples of 28 pixels.
Vision placeholders
4,608
8 two-frame groups × 576 per group
Spatial patch grid48 × 48 patches, merged 2 × 2
Odd-frame paddingNo frame is repeated.
02 / Temporal groups
Where the frames go
An odd final frame is repeated to complete the last pair. Each group receives the same placeholder count when processed dimensions are constant.
03 / How it is calculated
Three factors, one count
ceil(sampled frames ÷ 2)×(processed height ÷ 28)×(processed width ÷ 28)
Merge spatial patches
Each 14-pixel patch grid is merged 2 by 2. One placeholder covers a 28 by 28 processed-pixel region.
Count placeholders
The text processor repeats a vision placeholder for every merged spatial location in every temporal group.
What this number covers
This counts the repeated vision placeholders in the pinned MiniMax processor. It excludes timestamps, vision wrapper tokens, text, total prompt tokens, automatic frame sampling, resize decisions, model memory, speed and visual understanding. It does not require model weights or upload a video.
Pinned publisher sources
Inspect the implementation
All source links point to MiniMax M3.1 Flash revision 262ae6830934f3512211081b8a3f16425b93a239.