Mistral Large 4 vision encoder and projector

Vision tower and image-to-language connector extracted from Mistral Large 4, revision af26e7869af601441d5cdf75a1272469ddc3d059. This component does not include the language model.

The checkpoint preserves original tensor names, shapes and BF16 values: vision_encoder.*, pre_mm_projector_norm.*, patch_merger.* and vision_language_adapter.*. The 30-layer Pixtral-style vision tower has hidden size 2048, 16 attention heads, patch size 14, and axial 2-D rotary positions. The connector applies RMSNorm, learned 2×2 patch merging and a GELU MLP projecting to 6144 dimensions. The image processor configuration is retained in processor_config.json.

component.py supplies an independently loadable PyTorch wrapper based on the Pixtral implementation in Transformers 4.57.x. The wrapper maps original checkpoint names at loading time. Pass already preprocessed image tensors and their actual unpadded sizes; each size must be divisible by 28 for 2×2 patch merging. It does not contain a tokenizer, text generation, or the full official Mistral Large 4 implementation. See validation.json for actual numerical tests and their scope.

The model uses the Mistral Ecosystem Open-Weight License rather than Apache/MIT; the exact upstream license is included. Consult its terms for your use.