MoonViT-SO-400M
MoonViT-SO-400M is the image encoder extracted from Kimi VL Instruct. Moonshot initialized it from SigLIP-SO-400M and continued pretraining it before publishing the encoder separately.
Features instead of a written answer
The published example passes processed pixel values and image-grid dimensions into MoonViT and receives a list of image features. The separate package contains the vision implementation and weights, rather than the full Kimi VL language decoder.MoonViT-SO-400M model card
Its place in the original visual system
The original Kimi VL report identifies MoonViT as its native-resolution encoder. Its purpose is to represent visual detail for the downstream language model, including high-resolution inputs.Kimi VL: original technical report
Choose an output head for your task
Practical guidance. Use this encoder when building a system that consumes image features. Add a suitable projector or task head. Do not assume that raw features are already calibrated for a text-to-image search index.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | moonshotai/MoonViT-SO-400M |
| Role | Image-feature extraction; not a standalone chat model.MoonViT-SO-400M model card |
| Training ancestry | Continued pretraining of SigLIP-SO-400M.MoonViT-SO-400M model card |
| Weight license | MIT.MoonViT-SO-400M model card |
Related models
- Kimi-K2.5
Moonshot AI · Kimi K2
Multimodal K2 successor that combines image and video understanding with coding and reasoning.
Launch date Downloadable - Kimi-VL-A3B-Instruct
Moonshot AI · Kimi VL
Vision-language model for images, documents and video, built from Moonlight and the MoonViT encoder.
Launch date April 2025Downloadable