| --- |
| license: mit |
| library_name: kornia |
| base_model: moonshotai/Kimi-VL-A3B-Instruct |
| tags: |
| - kornia |
| - kimi-vl |
| - vision |
| - feature-extraction |
| --- |
| |
| # Kimi-VL-A3B-Instruct vision and projector weights for Kornia |
|
|
| > [!WARNING] |
| > This is an unofficial, temporary development checkpoint. It is hosted in a |
| > personal repository to support the implementation and review of |
| > [Kornia issue #3869](https://github.com/kornia/kornia/issues/3869). |
| > It is not an official Kornia model release, and its repository or location may |
| > change after the contribution is reviewed. |
|
|
| This repository contains the vision tower and multimodal projector weights from |
| [`moonshotai/Kimi-VL-A3B-Instruct`](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct), |
| converted to Kornia's `KimiVLModel` state-dictionary layout. |
|
|
| It does **not** contain the language-model weights or a complete Kimi-VL model. |
|
|
| ## Checkpoint |
|
|
| - File: `model.pt` |
| - Source: `moonshotai/Kimi-VL-A3B-Instruct` |
| - Contents: vision tower and multimodal projector only |
| - Format: PyTorch state dictionary |
| - Conversion: resizes the original 64 x 64 positional-embedding grid to the |
| current Kornia model's configured grid |
|
|
| ## License and attribution |
|
|
| The source model is published under the MIT license. See the |
| [official model repository](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct) |
| for its model card, license, usage information, and limitations. |
|
|
