From Pixels to Words -- Towards Native One-Vision Models at Scale

NEO-ov, from SenseTime Research and NTU’s S-Lab, is a fully native one-vision foundation model that removes the conventional vision encoder, projector, and post-hoc fusion stack used by modular VLMs such as Qwen3-VL and InternVL3.5. Instead, it serializes images, multi-image sets, and video frames directly into one decoder-only backbone with native patch embeddings, T-H-W decoupled attention, and Native RoPE, so pixel-word and pixel-pixel interactions begin at the earliest layers. The model is trained in three stages on 20M image-text pairs, nearly 60M multimodal samples, and 6M high-quality instruction samples, with context length expanded from 16K to 36K and visual resolution up to 4096^2; videos are sampled up to 128 frames. At 2B scale, NEO-ov reaches 54.7 on MMMU, 80.0 on MMBench, 91.2 on DocVQA, and 81.2 on OCRBench, exceeding prior native models and approaching modular counterparts. At 8B scale, it scores 68.1 on MMMU, 85.1 on MMBench, 91.9 on DocVQA, 81.6 on OCRBench, 67.4 on VideoMME, 70.7 on MVBench, and 90.0 on Mindcube, showing particularly strong gains in spatial intelligence, where it attains 64.8 on VSI-Bench and 90.0 on Mindcube. The main takeaway is that a monolithic, encoder-free architecture can scale to competitive multimodal reasoning while improving fine-grained spatial perception and cross-frame correspondence.