NeoMME: single-tower multimodal multilingual encoders beat sub-800M rivals on ViDoRe v3
CShorten30 · x · 2026-09-03
NeoMME is a family of 260M/800M multimodal-native multilingual bidirectional encoders that process text tokens and raw image patches in a single bidirectional Transformer — no pretrained vision tower, text encoder, or decoder — pretrained from scratch with a masked discrete-diffusion objective.
- 16,384-token context, enough to encode two 4K UHD images
- With joint dense + late-interaction heads, NeoMME-Retriever 260M hits 0.523 nDCG@10 on ViDoRe v3, beating all evaluated models below 800M; the 800M version reaches 0.556
- At matched 2048x2048 input on an NVIDIA L40S, the 260M model encodes pages with 2x the throughput of ColModernVBERT
- Hierarchical token pooling and asymmetric quantization compress late-interaction document embeddings for cheaper retrieval
More from Multimodal
- Jason Silva Made a 25-Minute Psychedelic AI Film in Five Days by Describing It Out Loud — lmoroney · 2026-10-21
- Ex-Brain Games host made a 25-min AI film in 5 days using 'vibe directing' in iMessage — lmoroney · 2026-10-21
- Motion Launches a General Video Agent That Edits Footage and Mimics Any Reference Style — mhdfaran · 2026-10-21
- LiFT Loops a Shared DiT Core: FID 3.34 Lower With 60% Fewer Parameters on ImageNet — burny_tech · 2026-10-21
- Ant open-sources 5 Ming-3.0 speech models spanning TTS, audio understanding and ASR — Med1_Ai · 2026-10-21
- What keeps AI videos watchable: hooks, consistency and a payoff — aftahi_ai · 2026-10-21