H Company Open-Sources NeoMME: 260M Multimodal Encoder Matches ColQwen2.5

H Company open-sourced NeoMME, a set of 260M/800M multilingual multimodal encoders trained jointly on text tokens and raw image patches from scratch. The 260M model matches ColQwen2.5's retrieval performance with 1/14 the parameters and supports OCR-free document retrieval.

2026-09-07 ~ 2026-09-07 · 2 related posts

• H Company's NeoMME: 260M encoder matches 3.75B ColQwen2.5 on ViDoRe v3 with 14x fewer params — tomaarsen · 2026-09-07
• H Company Releases NeoMME: 260M/800M Multilingual Text-Image Encoders With Day-0 Sentence Transformers Support — lateinteraction · 2026-09-07