A page at the compared resolution is a countable number of tokens. Images arrive as non-overlapping 32x32 patches projected by a small MLP [10], so a 2048x2048 page is 64 patches on a side and 4,096 patch tokens in total [1], about a quarter of the model's 16,384-token context [12][7]. That is the unit of work under the throughput figure. Multiply it out: 51 pages per second [5] is roughly 183,600 pages an hour on one card [2], and the doubling claim implies about 25.5 pages per second for ColModernVBERT in the same harness [3].
For those numbers to transfer you need their conditions: an L40S, a matched 2048x2048 input, whole page images [5]. NeoMME keeps native size and aspect ratio [11], which is the feature that makes throughput corpus-dependent. Denser pages spend more tokens. If your archive is 300 dpi A4 scans, recompute before sizing the fleet.
The storage figure is arithmetic worth checking. Roughly 1.5 MB down to 6 kB per page [6] is 250x if you count a megabyte as 1000 kB and 256x at 1024, so the quoted 255x sits inside that rounding either way [4]. The writeup credits the reduction to hierarchical token pooling and asymmetric quantization [6], and both are index-side techniques rather than properties of the backbone. My reading is that a fair comparison applies the same compression to a ColPali-style baseline before the ratio gets attributed to the architecture.
The architecture argument is a subtraction, and it is the part I find convincing. Retrieval, classification and token labeling never sample text, so a causal decoder is parameter and compute overhead you carry and never use for its purpose [9]. ModernVBERT went bidirectional on the text side but kept a separately pretrained SigLIP2 vision tower [8]. NeoMME drops both and trains a single bidirectional Transformer from scratch under a masked discrete-diffusion objective [2]. The training detail that makes that plausible is the corruption schedule: multimodal examples are masked at rates between 0.3 and 1 with the image patches left visible, so the model cannot recover the missing word from surrounding text and has to read the pixels [15]. Of about 524 billion packed tokens per model, 290 billion come from text-only examples [16], leaving roughly 234 billion, about 45 percent, in examples that carry an image [6].
Adoption cost is low: Apache 2.0 checkpoints, loadable in Hugging Face Transformers [7], 260M parameters at the small end [1]. The post also supplies a pronunciation guide, "nee-oh-me" [18], which is optimistic about how often the name will need saying aloud. The gap is evidence. Frontier placement on ViDoRe v3 for nDCG@10 against model size is asserted [4] without the scores appearing in the text supplied to us, and retention of more than 95 percent [6] means the 255x saving can cost as much as 5 percent of baseline nDCG@10 [8]. On a corpus where the right page routinely lands at rank 9, that is the number to measure first.
