Most systems that search inside scanned documents borrow an engine built for something else. H Company wants to replace that habit with purpose-built hardware-light models, and it has released two of them under the NeoMME name.
The 260M and 800M variants are bidirectional encoders that drop the two heaviest parts of the standard recipe: a separately trained vision tower and a causal decoder that never emits a token. Instead, one Transformer handles multilingual text and raw 32×32 image patches in the same layers, and the whole network is trained from a random start.
H Company’s argument is that mainstream retrievers descended from ColPali carry generation machinery they never use, which inflates both parameters and latency for a job that only needs representations. On the ViDoRe v3 retrieval benchmark, the 260M model’s fine-tuned retriever variant posts 0.523 nDCG@10.
Speed is where the small size shows. A single NVIDIA L40S indexes 51.3 pages per second with the 260M checkpoint, and a query resolves in 78.3 milliseconds even when the host is CPU-only. Text enters through an ALBERT-style factorized embedding, while images are split into non-overlapping 32×32 patches and projected by a two-layer MLP, with no patch-merging module and no SigLIP tower.
Licensing is Apache 2.0, and Hugging Face Transformers supports both checkpoints from day one. For teams indexing contracts, manuals and invoices on their own hardware, the pitch is a smaller footprint for the same retrieval class as far heavier open models.
