Kimi-VL-A3B-Instruct vision encoder and projector for Kornia

This repository hosts the vision tower (MoonViT) and multimodal projector weights of moonshotai/Kimi-VL-A3B-Instruct, repackaged into the state-dictionary layout of Kornia's KimiVLModel. It does not contain the language-model weights.

Fidelity

Every tensor value is bitwise identical to the original checkpoint (bf16, no resizing, no re-quantization). The only transformations are structural:

The fused attention wqkv weights/biases are sliced into separate q_proj / k_proj / v_proj tensors (pure slicing, values unchanged).
The learnable positional embedding is kept at its original 64×64 grid (4096 positions) and reshaped from (64, 64, 1152) to (1, 4096, 1152). Kornia's MoonViT bicubically interpolates it at runtime for other input resolutions, matching the upstream implementation's native-resolution design.
Keys are renamed to the Kornia module hierarchy (vision_tower.* -> vision_encoder.*, multi_modal_projector.* -> projector.*).

Provenance: converted from model-00001-of-00007.safetensors (sha256 a5ef3ebd9727f82e34417a778317d2cc9c08762fe0bc4a2ee333b8a52cf7c1a5) of moonshotai/Kimi-VL-A3B-Instruct. The conversion metadata is embedded in the safetensors header.

Usage

from kornia.models.kimi_vl import KimiVLBuilder

model = KimiVLBuilder.from_pretrained_hf().eval()

Files

model.safetensors — 443 tensors, 447,545,840 bf16 parameters (~895 MB). Safetensors only; no pickle files are hosted here.

License and attribution

The original model is published by Moonshot AI under the MIT license, which permits redistribution of original and modified versions with attribution. See the official model card and the Kimi-VL technical report for details, usage information, and limitations.

@misc{kimiteam2025kimivl,
  title={{Kimi-VL} Technical Report},
  author={Kimi Team},
  year={2025},
  eprint={2504.07491},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2504.07491},
}