model_1_ptm q8 v2 (template support)

The same qualified mixed-int8 model_1_ptm bundle as model1-ptm-q8-v1, re-exported to carry template_single_embedding and template_projection.

Those two tensors are siblings of the template module in AlphaFold rather than part of it, so v1 shipped the whole template pair stack without them. model_1_ptm sets embed_torsion_angles, so a prediction that uses a custom template needs them to embed the template's torsion angles into its MSA row.

339 tensors in 8 shards, 97.4 MiB. Quantized qualification against the float32 run is unchanged: mean pLDDT delta under 0.2, max atom error 1.64 A.

AlphaFold2 model 1 PTM q8 browser bundle

Qualified mixed q8 browser weights for AlphaFold2 model 1 PTM. Learned tensors use symmetric per-block int8 storage with float16 scales; structure and geometry tensors remain float32. Neural-network arithmetic is float32 after decoding. Includes immutable shard SHA-256 declarations and the official CC BY 4.0 weights license. Qualified against float32 for query-only and deep-MSA inputs in Chrome on NVIDIA GB10 and Apple M4 Pro. Archive SHA-256: 6f201ac7b90d3733b35566300c6f6ad3ebc38c39d2f0c1958eb3d2ad24de41be.

AlphaFold-Multimer-v3 model 1 f16 browser bundle

Qualified mixed-f16 browser weights for AlphaFold-Multimer-v3 model 1. Structure and geometry tensors remain float32, and neural-network arithmetic is float32 after decoding. Includes immutable shard SHA-256 declarations and the official CC BY 4.0 weights license. Qualified against official JAX and float32 WebGPU references for query-only and paired/unpaired inputs in Chrome on NVIDIA GB10 and Apple M4 Pro. Archive SHA-256: 47fd4b67f0dc824c6d9199f7da71d37f4ef41b495b83ca658e4042817f44e8c1.

AlphaFold2 model 1 PTM browser bundle

Reduced float32 browser inference bundle containing only the 335 tensors required by AlphaFold2 WebGPU.