Visual document retrieval has long suffered from architectural bloat. Most production systems rely on repurposed generative vision-language models, carrying heavy vision towers and causal decoders that sit idle during encoding tasks. H Company is challenging this status quo with NeoMME, a family of bidirectional encoders that strips away the excess. By processing multilingual text and raw 32×32 RGB image patches through a single Transformer trained from scratch, NeoMME eliminates the parameter overhead of separate vision components. The result is a leaner, faster system designed specifically for representation learning rather than token generation.
The performance metrics are striking for such a compact architecture. The 260M-parameter variant achieves a 0.523 nDCG@10 score on the ViDoRe v3 benchmark. This places it within 0.002 points of ColQwen2.5-v0.2, a model with 3.75 billion parameters. In practical terms, NeoMME is 14.4 times smaller while delivering comparable accuracy on visual tasks. It also outperforms other sub-300M models by a wide margin, beating the next best competitor by 26.1 points. However, the model isn’t perfect across all domains. Text-only retrieval on BEIR-15 lags behind specialized models like LateOn, likely due to seeing only 430K text query examples during training compared to the hundreds of millions used by rivals.
Efficiency extends beyond parameter count to actual deployment speed and storage. On a single NVIDIA L40S GPU, NeoMME-260M indexes 51.3 pages per second, nearly doubling the throughput of ColModernVBERT. Query encoding takes just 78.3 ms on a CPU-only host, making it viable for edge devices or cost-sensitive cloud environments. Storage costs are also drastically reduced through hierarchical token pooling and asymmetric quantization. These techniques shrink the index size from roughly 1.5 MB per page to just 6 kB, a 255.5x reduction that retains over 95% of baseline accuracy. With Apache 2.0 licensing and immediate Hugging Face support, H Company has lowered the barrier for high-performance visual search.
