H Company introduced NeoMME on Sept. 3, 2026, as 260M and 800M multilingual multimodal encoders for representation and visual-document retrieval, not generative vision-language modeling. AI Breaking Wire reported the same release date and sizes the same day.
NeoMME routes text tokens and raw 32x32 image patches through one bidirectional Transformer, without a separately pretrained vision tower or causal decoder. Both sizes support a 16,384-token context. Hugging Face Transformers documentation independently describes that shared-encoder design.
NeoMME-Retriever is fine-tuned for visual document retrieval and returns dense and late-interaction embeddings in one forward pass. On the model card's reported ViDoRe v3 results, late-interaction nDCG@10 is 0.5226 for the 260M checkpoint and 0.5560 for the 800M checkpoint. H Company says the 260M retriever is within 0.002 nDCG@10 of ColQwen2.5 while using about 14 times fewer parameters; those comparisons remain company-reported.
At a matched 2048x2048 image input on one NVIDIA L40S, H Company reports about 51 pages per second for the 260M retriever, roughly twice ColModernVBERT's throughput in the same write-up. Hierarchical token pooling and asymmetric quantization, the company says, cut late-interaction index storage from about 1.5 MB to 6 kB per page, a 255x reduction, while retaining more than 95% of baseline nDCG@10; a less aggressive setting reaches 39 kB with more than 99% retention.
Training is from scratch with an image-conditioned masked discrete-diffusion text objective. The technical report and primary post say each model processes about 524 billion packed input tokens, including 290 billion text-only tokens. Checkpoints are Apache 2.0, and NeoMME was contributed to Hugging Face Transformers on Aug. 31, 2026.
What remains open is independent reproduction of the ViDoRe, throughput, and compression figures, plus language-by-language quality, production readiness, and how dense versus late-interaction and compression settings trade off in live serving. KiaDev's summary notes that release text alone does not supply operational measurements sufficient to pick a production checkpoint. Company benches and independent coverage should be weighed separately until those points are settled.