remove chai1-int5 and esm2-3b-int8: the CUDA ports read af3-any-model's own chai1 and ESM2 blobs (sokrypton/af3-any-model 28141c7), and the page offers neither

esm2-3b-int8: ESM2 3B (MIT), int8 per row, q|k|v stacked - Chai-1's token features

chai1-int5: Chai-1 (Apache-2.0), int5 group 32, with chai-lab's structure token-pair weights

af2 monomer: the template single features (template_single_embedding, template_projection) - model_1 and model_2 as a new float32 shard, models 3-5 absent; existing shards unchanged

ESMFold2 300M with Synthyra's confidence head (int5)

ESMFold2 600M with Synthyra's confidence head (int5)

AlphaFold 2 multimer model_5 as a delta

AlphaFold 2 multimer model_4 as a delta

AlphaFold 2 multimer model_3 as a delta

AlphaFold 2 multimer model_2 as a delta

AlphaFold 2 at int5: af2-monomer-5

AlphaFold 2 at int5: af2-monomer-4

AlphaFold 2 at int5: af2-monomer-3

AlphaFold 2 at int5: af2-monomer-2

AlphaFold 2 at int5: af2-multimer-int5

AlphaFold 2 at int5: af2-monomer-int5

Upload folder using huggingface_hub

Upload folder using huggingface_hub

opendde-int5: re-export - 833-channel single conditioning (PADDED_SINGLE_COND) and eight encoder tensors that were zero

protenix2-int5: Protenix-v2 whole at int5 group 32, 8 shards

boltz2-int5: Boltz-2 whole (trunk, diffusion, confidence) at int5 group 32, 8 shards

Upload folder using huggingface_hub

Correct AF2's distogram bin breaks: 2.3125 / 21.6875, not 2 / 22

Eight bundles, balanced shards, no dead copies