ProGen3

Profluent's autoregressive protein language model wrapper (proteingen.models.ProGen3).

Unlike masked models (ESMC, ESM3, DPLM-2), ProGen3 generates left-to-right (N→C terminal):

• Output dim: 134
• Embedding dim: checkpoint-dependent (EMB_DIM read from weights)
• Primary use cases: autoregressive generation, sequence scoring, embedding extraction

Available checkpoints

Checkpoint	Params
Profluent-Bio/progen3-112m	112M
Profluent-Bio/progen3-219m	219M
Profluent-Bio/progen3-339m	339M
Profluent-Bio/progen3-762m	762M
Profluent-Bio/progen3-1b	1B
Profluent-Bio/progen3-3b	3B

from proteingen.models import ProGen3

model = ProGen3("Profluent-Bio/progen3-112m").cuda()

Generation APIs

Fixed-length generation via sample()

from proteingen.sampling import sample

init_x = ["<mask>" * 100 for _ in range(4)]
traj = sample(model, init_x, in_order="left_to_right")
print(traj["sequences"])

Open-ended generation via generate()

result = model.generate(n=4, max_new_tokens=256, temperature=0.8, top_p=0.95)
print(result["sequences"])

Sequence scoring

scores = model.score(["ACDEFGHIKLMNPQRSTVWY"])
print(scores["log_likelihood"], scores["perplexity"])

Dependency caveat

The upstream progen3 package may need additional runtime deps (not declared in package metadata), notably megablocks/grouped_gemm (and often Flash Attention) for MoE checkpoints.

See also: Autoregressive Generation (ProGen3)