NeoMME: H Company Ships Multilingual Multimodal Encoders in Transformers

What shipped

H Company released NeoMME on September 3, 2026 — a family of 260M and 800M multilingual multimodal encoders under the Apache 2.0 license, available now in Hugging Face Transformers with checkpoints published. No pretrained vision tower, no causal decoder: one bidirectional Transformer processes text tokens and raw image patches, trained from scratch with a masked discrete-diffusion objective.

What changed

NeoMME-Retriever, the fine-tuned variant for visual document retrieval (ColPali's page-image approach), returns dense and late-interaction embeddings in a single forward pass. Both sizes sit on the ViDoRe v3 Pareto frontier for nDCG@10 vs model size. At matched 2048×2048 input on an L40S, the 260M model encodes about 51 pages/second — roughly twice ModernVBERT's throughput. Hierarchical token pooling plus asymmetric quantization shrinks late-interaction index storage from ~1.5 MB to ~6 kB per page (255×) while keeping over 95% of baseline nDCG@10.

Why a builder cares

Retrieval and classification workloads don't generate text, so they shouldn't pay for a decoder. NeoMME gives you a self-contained multimodal encoder at half the overhead of VLM-derived retrievers, with a 255× smaller index that makes late-interaction retrieval practical at scale — and it's Apache 2.0.