This file is generated from src/fastplms/models.toml. A listed capability is
selectable. Strict-parity exceptions are documented in the checkpoint cards.
| Family |
Architecture |
Checkpoints |
Public input |
AutoClasses |
Tokenizer class |
esm2 |
ESM2 |
5 |
Amino-acid sequences tokenized to residue IDs |
AutoConfig, AutoModel, AutoModelForMaskedLM, AutoModelForSequenceClassification, AutoModelForTokenClassification |
n/a |
esm_plusplus |
ESMC |
3 |
Amino-acid sequences tokenized to residue IDs |
AutoConfig, AutoModel, AutoModelForMaskedLM |
n/a |
esm3 |
ESM3 |
1 |
Sequence, structure, and function tracks prepared through the multimodal helpers |
AutoConfig, AutoModel |
n/a |
e1 |
E1 |
3 |
Raw amino-acid sequences prepared by the native E1 adapter |
AutoConfig, AutoModel, AutoModelForMaskedLM, AutoModelForSequenceClassification, AutoModelForTokenClassification |
n/a |
dplm |
DPLM |
3 |
Amino-acid sequences tokenized to masked or partially masked residue IDs |
AutoConfig, AutoModel, AutoModelForMaskedLM, AutoModelForSequenceClassification, AutoModelForTokenClassification |
n/a |
dplm2 |
DPLM2 |
3 |
Tokenized amino-acid and structure tracks with explicit modality boundaries |
AutoConfig, AutoModel, AutoModelForMaskedLM, AutoModelForSequenceClassification, AutoModelForTokenClassification |
fastplms.models.dplm2.tokenization_dplm2.DPLM2Tokenizer |
ankh |
ANKH |
5 |
Amino-acid sequences tokenized for encoder or sequence-to-sequence use |
AutoConfig, AutoModel, AutoModelForMaskedLM, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification, AutoModelForTokenClassification |
n/a |
boltz2 |
Boltz2 |
1 |
Raw amino-acid sequences through the convenience API, or prepared model features |
AutoConfig, AutoModel |
n/a |
esmfold |
ESMFold |
1 |
Raw amino-acid sequences through folding helpers, or prepared residue tensors |
AutoConfig, AutoModel |
n/a |
esmfold2 |
ESMFold2 |
4 |
Raw amino-acid sequences or typed molecular-complex specifications; low-level forward accepts prepared feature tensors |
AutoConfig, AutoModel |
n/a |
pretrained means the advertised head is present in the checkpoint. base weights + untrained task head means the task head must be trained before use. FastPLMs extension is an integration or head that is not an official pretrained ANKH capability.
| Family |
AutoClass |
Weight status |
esm2 |
AutoConfig |
FastPLMs extension |
esm2 |
AutoModel |
pretrained |
esm2 |
AutoModelForMaskedLM |
pretrained |
esm2 |
AutoModelForSequenceClassification |
base weights + untrained task head |
esm2 |
AutoModelForTokenClassification |
base weights + untrained task head |
esm_plusplus |
AutoConfig |
FastPLMs extension |
esm_plusplus |
AutoModel |
pretrained |
esm_plusplus |
AutoModelForMaskedLM |
pretrained |
esm3 |
AutoConfig |
FastPLMs extension |
esm3 |
AutoModel |
pretrained |
e1 |
AutoConfig |
FastPLMs extension |
e1 |
AutoModel |
pretrained |
e1 |
AutoModelForMaskedLM |
pretrained |
e1 |
AutoModelForSequenceClassification |
base weights + untrained task head |
e1 |
AutoModelForTokenClassification |
base weights + untrained task head |
dplm |
AutoConfig |
FastPLMs extension |
dplm |
AutoModel |
pretrained |
dplm |
AutoModelForMaskedLM |
pretrained |
dplm |
AutoModelForSequenceClassification |
base weights + untrained task head |
dplm |
AutoModelForTokenClassification |
base weights + untrained task head |
dplm2 |
AutoConfig |
FastPLMs extension |
dplm2 |
AutoModel |
pretrained |
dplm2 |
AutoModelForMaskedLM |
pretrained |
dplm2 |
AutoModelForSequenceClassification |
base weights + untrained task head |
dplm2 |
AutoModelForTokenClassification |
base weights + untrained task head |
ankh |
AutoConfig |
FastPLMs extension |
ankh |
AutoModel |
pretrained |
ankh |
AutoModelForMaskedLM |
FastPLMs extension |
ankh |
AutoModelForSeq2SeqLM |
pretrained |
ankh |
AutoModelForSequenceClassification |
base weights + untrained task head |
ankh |
AutoModelForTokenClassification |
base weights + untrained task head |
boltz2 |
AutoConfig |
FastPLMs extension |
boltz2 |
AutoModel |
pretrained |
esmfold |
AutoConfig |
FastPLMs extension |
esmfold |
AutoModel |
pretrained |
esmfold2 |
AutoConfig |
FastPLMs extension |
esmfold2 |
AutoModel |
pretrained |
| Family |
Attention |
Precision |
BF16 execution |
Extra |
Reference |
esm2 |
eager, sdpa, flex_attention, flash_attention_2, flash_attention_3 |
default |
fp32_parameters_autocast |
core |
reference-esm2 |
esm_plusplus |
eager, sdpa, flex_attention, flash_attention_2, flash_attention_3 |
default, fp8 (experimental) |
static_parameters |
core |
reference-biohub-esm |
esm3 |
eager, sdpa, flex_attention |
default |
fp32_parameters_autocast |
core |
reference-biohub-esm |
e1 |
sdpa, flex_attention |
default |
static_parameters |
core |
reference-e1 |
dplm |
eager, sdpa, flex_attention, flash_attention_3 |
default |
fp32_parameters_autocast |
core |
reference-dplm |
dplm2 |
sdpa |
default |
fp32_parameters_autocast |
core |
reference-dplm |
ankh |
eager, sdpa |
default |
static_parameters |
core |
reference-ankh |
boltz2 |
eager |
default |
fp32_parameters_autocast |
structure |
reference-boltz2 |
esmfold |
eager, sdpa, flex_attention |
default |
fp32_parameters_autocast |
structure |
reference-esmfold |
esmfold2 |
eager, sdpa, flex_attention |
auto, fp32, bf16, fp8 (experimental) |
fp32_parameters_autocast |
structure |
reference-esmfold2 |
| Family |
Checkpoint terms |
Hub license |
Weight publication |
Tiers |
esm2 |
MIT |
mit |
manifest policy |
check, compliance, feature, artifact, benchmark |
esm_plusplus |
MIT |
mit |
manifest policy |
check, compliance, feature, artifact, benchmark |
esm3 |
MIT |
mit |
manifest policy |
check, compliance, feature, artifact, benchmark |
e1 |
Profluent-E1-Agreement |
other (Profluent-E1 Clickthrough License Agreement) |
manifest policy |
check, compliance, feature, artifact, benchmark |
dplm |
Apache-2.0 |
apache-2.0 |
manifest policy |
check, compliance, feature, artifact, benchmark |
dplm2 |
Apache-2.0 |
apache-2.0 |
manifest policy |
check, compliance, feature, artifact, benchmark |
ankh |
CC-BY-NC-SA-4.0 |
cc-by-nc-sa-4.0 |
manifest policy |
check, compliance, feature, artifact, benchmark |
boltz2 |
MIT |
mit |
manifest policy |
structure, artifact, benchmark |
esmfold |
MIT |
mit |
manifest policy |
check, compliance, structure, feature, artifact, benchmark |
esmfold2 |
MIT |
mit |
manifest policy |
check, compliance, structure, feature, artifact, benchmark |
| ID |
Family |
Repository |
Path |
SHA-256 |
Size |
License |
Trust boundary |
Offline behavior |
esmfold2_ccd |
esmfold2 |
biohub/ESMFold2 |
ccd.pkl |
9ff44b1927c6b9198e38ffe0928706827a09a350c15530beeeabebfa88038fc5 |
417306584 |
MIT |
hash_pinned_pickle |
requires_cached_verified_file |
A nonzero unresolved-file count blocks release. It is not permission to
omit that file from checkpoint, tokenizer, artifact, or compliance checks.