The models coresample's page runs (https://github.com/davidackerman/coresample)

Each file here is a published model converted to ONNX by web/scripts/export_model.py, its
larger weights stored as float16 (rounded from float32, run in float32). The models are their
authors' work, under their own licences (not coresample's BSD-3-Clause).

dinov2-small-112.onnx, dinov2-small-224.onnx, dinov2-small-112-tokens.onnx
  DINOv2 ViT-S/14: https://huggingface.co/facebook/dinov2-small
  (https://github.com/facebookresearch/dinov2)
  Copyright (c) Meta Platforms, Inc. and affiliates. Apache License 2.0: LICENSE-Apache-2.0.txt
  Changed: converted to ONNX at fixed input sizes (112 and 224 px); dinov2-small-112 also pools
  its outputs into three unit-length 384-d parts (class token, mean and centre-weighted mean of
  the patch tokens).

pecore-t16-112-tokens.onnx
  Perception Encoder PE-Core-T16-384: https://huggingface.co/facebook/PE-Core-T16-384
  (https://github.com/facebookresearch/perception_models), loaded through timm
  (https://huggingface.co/timm/vit_pe_core_tiny_patch16_384.fb, Apache License 2.0)
  Copyright (c) Meta Platforms, Inc. and affiliates. Apache License 2.0: LICENSE-Apache-2.0.txt
  Changed: converted to ONNX at 112 px input (timm resizes its position embeddings to the 7 x 7
  grid); outputs the last layer's tokens, each 192-d token followed by 192 zeros (384-d).

morphem-112-tokens.onnx
  MorphEm ViT-S/16: https://huggingface.co/CaicedoLab/MorphEm
  (https://github.com/CaicedoLab/CHAMMI-75), by Vidit Agrawal, John Peters and Juan Caicedo
  Copyright (c) 2025 CaicedoLab. MIT License: LICENSE-MIT-MorphEm.txt
  Changed: loaded into timm's ViT-S/16 with one input channel, the DINO training head left out,
  position embeddings resized to the 7 x 7 grid of a 112 px input; converted to ONNX; outputs the
  last layer's tokens.