Model support

August 5, 2026 · View on GitHub

This file is generated from src/fastplms/models.toml. A listed capability is selectable. Strict-parity exceptions are documented in the checkpoint cards.

Family interfaces

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`, `AutoModelForSequenceClassification`, `AutoModelForTokenClassification`	`n/a`
`esm3`	ESM3	1	Sequence, structure, and function tracks prepared through the multimodal helpers	`AutoConfig`, `AutoModel`, `AutoModelForSequenceClassification`, `AutoModelForTokenClassification`	`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`, `AutoModelForSequenceClassification`, `AutoModelForTokenClassification`	`n/a`
`esmfold2`	ESMFold2	4	Raw amino-acid sequences or typed molecular-complex specifications; low-level forward accepts prepared feature tensors	`AutoConfig`, `AutoModel`, `AutoModelForSequenceClassification`, `AutoModelForTokenClassification`	`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`
`esm_plusplus`	`AutoModelForSequenceClassification`	`base weights + untrained task head`
`esm_plusplus`	`AutoModelForTokenClassification`	`base weights + untrained task head`
`esm3`	`AutoConfig`	`FastPLMs extension`
`esm3`	`AutoModel`	`pretrained`
`esm3`	`AutoModelForSequenceClassification`	`base weights + untrained task head`
`esm3`	`AutoModelForTokenClassification`	`base weights + untrained task head`
`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`
`esmfold`	`AutoModelForSequenceClassification`	`base weights + untrained task head`
`esmfold`	`AutoModelForTokenClassification`	`base weights + untrained task head`
`esmfold2`	`AutoConfig`	`FastPLMs extension`
`esmfold2`	`AutoModel`	`pretrained`
`esmfold2`	`AutoModelForSequenceClassification`	`base weights + untrained task head`
`esmfold2`	`AutoModelForTokenClassification`	`base weights + untrained task head`

Family execution

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 release contracts

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`

Runtime assets

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`

Checkpoints

ID	Family	Size	FastPLMs checkpoint	Official checkpoint	Artifact source	State transform	Generation contract	MSA conditioning
`esm2_8m`	`esm2`	`small`	Synthyra/ESM2-8M	facebook/esm2_t6_8M_UR50D	`fast`	`esm2_hf_to_fastplms_v1`	`not_applicable`	not applicable
`esm2_35m`	`esm2`	`small`	Synthyra/ESM2-35M	facebook/esm2_t12_35M_UR50D	`fast`	`esm2_hf_to_fastplms_v1`	`not_applicable`	not applicable
`esm2_150m`	`esm2`	`medium`	Synthyra/ESM2-150M	facebook/esm2_t30_150M_UR50D	`fast`	`esm2_hf_to_fastplms_v1`	`not_applicable`	not applicable
`esm2_650m`	`esm2`	`large`	Synthyra/ESM2-650M	facebook/esm2_t33_650M_UR50D	`fast`	`esm2_hf_to_fastplms_v1`	`not_applicable`	not applicable
`esm2_3b`	`esm2`	`xlarge`	Synthyra/ESM2-3B	facebook/esm2_t36_3B_UR50D	`fast`	`esm2_hf_to_fastplms_v1`	`not_applicable`	not applicable
`esmc_small`	`esm_plusplus`	`medium`	Synthyra/ESMplusplus_small	biohub/ESMC-300M	`fast`	`esmc_to_fastplms_v1`	`not_applicable`	not applicable
`esmc_large`	`esm_plusplus`	`large`	Synthyra/ESMplusplus_large	biohub/ESMC-600M	`fast`	`esmc_to_fastplms_v1`	`not_applicable`	not applicable
`esmc_6b`	`esm_plusplus`	`xlarge`	Synthyra/ESMplusplus_6B	biohub/ESMC-6B	`fast`	`esmc_to_fastplms_v1`	`not_applicable`	not applicable
`esm3_small`	`esm3`	`large`	Synthyra/ESM3_small	biohub/esm3-sm-open-v1	`fast`	`esm3_to_fastplms_v1`	`not_applicable`	not applicable
`e1_150m`	`e1`	`small`	Synthyra/Profluent-E1-150M	Profluent-Bio/E1-150m	`fast`	`e1_to_fastplms_v1`	`not_applicable`	not applicable
`e1_300m`	`e1`	`medium`	Synthyra/Profluent-E1-300M	Profluent-Bio/E1-300m	`fast`	`e1_to_fastplms_v1`	`not_applicable`	not applicable
`e1_600m`	`e1`	`large`	Synthyra/Profluent-E1-600M	Profluent-Bio/E1-600m	`fast`	`e1_to_fastplms_v1`	`not_applicable`	not applicable
`dplm_150m`	`dplm`	`small`	Synthyra/DPLM-150M	airkingbd/dplm_150m	`fast`	`dplm_to_fastplms_v1`	`required`	not applicable
`dplm_650m`	`dplm`	`large`	Synthyra/DPLM-650M	airkingbd/dplm_650m	`fast`	`dplm_to_fastplms_v1`	`required`	not applicable
`dplm_3b`	`dplm`	`xlarge`	Synthyra/DPLM-3B	airkingbd/dplm_3b	`fast`	`dplm_to_fastplms_v1`	`required`	not applicable
`dplm2_150m`	`dplm2`	`small`	Synthyra/DPLM2-150M	airkingbd/dplm2_150m	`official`	`dplm2_to_fastplms_v1`	`required`	not applicable
`dplm2_650m`	`dplm2`	`large`	Synthyra/DPLM2-650M	airkingbd/dplm2_650m	`official`	`dplm2_to_fastplms_v1`	`required`	not applicable
`dplm2_3b`	`dplm2`	`xlarge`	Synthyra/DPLM2-3B	airkingbd/dplm2_3b	`official`	`dplm2_to_fastplms_v1`	`official_unavailable`	not applicable
`ankh_base`	`ankh`	`medium`	Synthyra/ANKH_base	ElnaggarLab/ankh-base	`official`	`ankh_t5_to_fastplms_v1`	`required`	not applicable
`ankh_large`	`ankh`	`large`	Synthyra/ANKH_large	ElnaggarLab/ankh-large	`official`	`ankh_t5_to_fastplms_v1`	`required`	not applicable
`ankh2_large`	`ankh`	`large`	Synthyra/ANKH2_large	ElnaggarLab/ankh2-ext2	`official`	`ankh_t5_to_fastplms_v1`	`required`	not applicable
`ankh3_large`	`ankh`	`large`	Synthyra/ANKH3_large	ElnaggarLab/ankh3-large	`official`	`ankh_t5_to_fastplms_v1`	`required`	not applicable
`ankh3_xl`	`ankh`	`xlarge`	Synthyra/ANKH3_xl	ElnaggarLab/ankh3-xl	`official`	`ankh_t5_to_fastplms_v1`	`required`	not applicable
`boltz2`	`boltz2`	`structure`	Synthyra/Boltz2	boltz-community/boltz-2	`fast`	`boltz2_inference_core_v1`	`not_applicable`	not applicable
`esmfold`	`esmfold`	`structure`	Synthyra/FastESMFold	facebook/esmfold_v1	`fast`	`esmfold_meta_to_fastplms_v1`	`not_applicable`	not applicable
`esmfold2`	`esmfold2`	`structure`	Synthyra/ESMFold2	biohub/ESMFold2	`fast`	`identity`	`not_applicable`	`optional` (full checkpoint)
`esmfold2_fast`	`esmfold2`	`structure`	Synthyra/ESMFold2-Fast	biohub/ESMFold2-Fast	`fast`	`identity`	`not_applicable`	`none` (Fast; MSA inputs rejected)
`esmfold2_experimental_cutoff2025`	`esmfold2`	`structure`	Synthyra/ESMFold2-Experimental-Cutoff2025	biohub/ESMFold2-Experimental-Cutoff2025	`fast`	`identity`	`not_applicable`	`optional` (full checkpoint)
`esmfold2_experimental_fast_cutoff2025`	`esmfold2`	`structure`	Synthyra/ESMFold2-Experimental-Fast-Cutoff2025	biohub/ESMFold2-Experimental-Fast-Cutoff2025	`fast`	`identity`	`not_applicable`	`none` (Fast; MSA inputs rejected)

A nonzero unresolved-file count blocks release. It is not permission to omit that file from checkpoint, tokenizer, artifact, or compliance checks.

Family interfaces

AutoClass weight status

Family execution

Family release contracts

Runtime assets

Checkpoints