AI Signal 130
NeoMME introduces efficient multimodal-native multilingual encoder without separate vision tower
↗
NeoMME provides a single bidirectional Transformer that processes text and image patches from scratch, enabling efficient multimodal encoding without separate vision or language towers.
Engineers can achieve higher throughput, with the 260M model encoding about 51 pages per second on an NVIDIA L40S GPU, roughly twice the speed of ColModernVBERT at the same resolution. Hierarchical token pooling and asymmetric quantization reduce storage per page from roughly 1.5 MB to about 6 kB while preserving over 95% of baseline nDCG@10. The model’s placement on the ViDoRe v3 Pareto frontier and its availability under the Apache 2.0 license in Hugging Face Transformers make it a practical choice for visual document retrieval pipelines.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
NeoMME uses a single bidirectional Transformer for text tokens and image patches, trained from scratch with a masked discrete-diffusion objective.
The 260M variant processes approximately 51 pages per second on an NVIDIA L40S GPU at 2048×2048 input, about twice the throughput of ColModernVBERT.
Hierarchical token pooling and asymmetric quantization cut late-interaction index storage from ~1.5 MB to ~6 kB per page (≈255× smaller) while retaining >95% of baseline nDCG@10.
THE CLUSTER