As of 2026, many visual document retrieval systems rely on repurposed generative models. ColPali, for example, adapts a vision-language model for retrieval tasks, but still carries a separate pretrained vision tower and a causal decoder that never actually generates tokens. This introduces unnecessary parameter and compute overhead for a task that only needs representations.
H Company has addressed this with NeoMME, a family of 260M and 800M bidirectional encoders that strips away both components. A single Transformer processes multilingual text tokens and raw 32Γ32 RGB image patches through shared layers, trained from random initialization. The retrieval-optimized model, NeoMME-Retriever, achieves 0.523 nDCG@10 on ViDoRe v3 at just 260M parameters.
Deployment Readiness
NeoMME is built for practical adoption. All checkpoints are open-sourced under Apache 2.0 and come with day-zero support in Hugging Face Transformers. The 260M model indexes 51.3 pages per second on a single NVIDIA GPU (2026 hardware reference), making it a strong candidate for high-throughput document retrieval pipelines.
Architectural Innovation
By removing the vision tower and causal decoder, NeoMME reduces model complexity while maintaining competitive retrieval accuracy. This design choice enables faster inference and lower memory consumption, particularly valuable for edge deployments and large-scale indexing workloads.
Next Steps
With the release, H Company provides model weights, integration guides, and benchmarking scripts. Developers can adopt NeoMME for tasks ranging from RAG pipelines to enterprise search. As multimodal retrieval evolves, this single-tower approach could become a new baseline in the field.
via MarkTechPost