ESMC

ESM-C masked language model, available in 300m and 600m parameter variants. Wraps the ESM model as a GenerativeModelWithEmbedding, providing both generative sampling and differentiable embedding extraction (used by LinearProbe and TAG).

• Output dim: 64 (33 real vocab + 31 alignment padding — handled automatically by MaskedModelLogitFormatter)
• Embedding dim: 960 (300m) — set dynamically from model weights
• LoRA support: yes, via apply_lora()

from proteingen.models import ESMC

model = ESMC("esmc_300m").cuda()  # or "esmc_600m"