Moonshot AI Opens MoonViT-SO-400M to Process Any Image Without Resizing

Most vision encoders make a quiet but costly assumption: every image should be the same size. Resize to 224x224, pad to a square, crop the edges -- whatever it takes to fit a fixed grid. MoonViT-SO-400M throws that assumption out. Moonshot AI has released the vision encoder from their Kimi-VL model as a standalone component under an MIT license, letting anyone drop a production-grade, native-resolution image encoder into their own multimodal stack.

The fixed-resolution tax

Standard Vision Transformers (ViTs) -- CLIP, SigLIP, the usual suspects -- were designed for fixed-resolution square inputs. A standard ViT requires every image to be the same size, such as 224x224 or 384x384. To accommodate different aspect ratios you either stretch the image, which distorts content, or pad with zeros, which wastes compute. For most tasks this is fine. For OCR on a scanned document, a 4K GUI screenshot, or a portrait photo with a 9:16 aspect ratio, it's a real problem -- you're throwing away exactly the information you need.

MoonViT is designed to natively process images at their varying resolutions, eliminating the need for complex sub-image splitting and splicing operations. It incorporates the packing method from NaViT, where images are divided into patches, flattened, and sequentially concatenated into 1D sequences. The result is that a 1920x1080 screenshot and a 400x600 portrait can sit in the same training batch without any resizing or padding.

Three things that make it different

• No forced resizing. MoonViT tokenizes each image at its native (H, W) and discards nothing -- a 1920x1080 screenshot and a 400x600 portrait can coexist in the same batch without any resizing.
• Zero wasted tokens. In vision, padding can mean more than 50% of the tokens in a batch are meaningless pad values. MoonViT packs all images end-to-end into one flat sequence and uses cu_seqlens to mark boundaries, so every token that enters the Transformer carries real signal.
• Dual positional encoding. The original SigLIP absolute positional embeddings become increasingly inadequate as image resolution increases. To address this, MoonViT incorporates 2D rotary positional embedding (RoPE) across the height and width dimensions, which improves the representation of fine-grained positional information, especially in high-resolution images. The absolute embed preserves the pretrained prior; the 2D RoPE handles resolution generalization.