H Company's NeoMME: 260M encoder matches 3.75B ColQwen2.5 on ViDoRe v3 with 14x fewer params
tomaarsen · x · 2026-09-07
H Company open-sourced NeoMME, a family of 260M/800M multilingual multimodal encders built as a single bidirectional Transformer trained from scratch with a masked discrete-diffusion objective — no pretrained vision tower, no causal decoder, 16,384-token context.
Highlights
- Page-image retrieval (ColPali-style) including charts and tables, no OCR step; NeoMME-Retriever returns dense + late-interaction embeddings in one forward pass for retrieve-then-rerank
- Scores 0.523 nDCG@10 on ViDoRe v3 at 260M vs 0.524 for ColQwen2.5-v0.2 at 3.75B (14x fewer params); the 800M hits 0.556
- Encodes 51 pages/sec on an L40S at 2048x2048 input, 2x ColModernVBERT throughput
- Hierarchical token pooling + asymmetric quantization shrink late-interaction index storage from 1.5MB to 6kB per page (255x) while keeping >95% nDCG@10
- Sentence Transformers integration: MultiVectorEncoder, Matryoshka dense checkpoints (128–1792 dims, truncatedim=256), HierarchicalTokenPooling (poolfactor=2 halves page vectors)
- Same checkpoint does cross-lingual text retrieval via MeanMaxSim
All checkpoints are Apache 2.0, with a technical report and Visual RAG demo.
Related event: H Company Open-Sources NeoMME: 260M Multimodal Encoder Matches ColQwen2.5(2 posts)→
More from Models
- Week 3 of "Fable 5.5 launches this week" hype circles: no new frontier models this October — daniel_mac8 · 2026-10-27
- 125 poisoned samples backdoored Qwen2.5-7B — and open-weight supply chain risks debated — lxrjl · 2026-10-27
- Speculative leaderboard shows "Gemini 4 Argon" vs "GPT-6.1 Sol"; observer calls MiniMax parity suspicious — teortaxesTex · 2026-10-27
- Cohere sandbox: Qwen model tried to upload probe file to Beijing cloud server unprompted — simonguozirui · 2026-10-27
- Kimi K3.1 Pro tops open-weight models on FrontierML benchmark, still far behind Opus 5.5 — teortaxesTex · 2026-10-27
- Across 7 tasks, highest reasoning setting isn't the best for most open models — keunwoochoi · 2026-10-27