AI for Bio · Model

AMPLIFY

Small protein language models arguing data quality matters more than scale

ProteinsMixedCanada, United StatesMaintained only中文

Overview

NameAMPLIFY · aliases: AMPLIFY, AMPLIFY 350M, AMPLIFY 120M
Full nameThe name is a model codename with no expanded form in the paper
Ambiguous names (not auto-linked)Amplify
FamilyNot a family entry. 120M and 350M are two sizes of one model, not separate members
OrganizationMila, Polytechnique Montréal, Amgen
Organization typeMixed · A joint project of an academic institute and a university lab with a large biopharma company's research arm
CountryCanada, United States
Model classProtein language model (small scale)
DomainProteins
First release · preprint
Basis: Preprint v1 posted; the 2024-09-19 first commits and 2024-09-30 weights do not count as first release source
Other milestones
  • First public evidence: · Preprint v1 source
  • Preprint v1: · Preprint v1 source
  • Code release: † · First repository commits source
  • Weights release: † · Weights published to a public model hub source
Latest update · paper revision
Basis: Preprint v2 revision
Paper revised to v2 (v1 was 2024-09-23); no journal version, and the repository and model hub carry only minor changes with the last default-branch commit on 2025-07-17 source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Only a paper revision and minor code changes, no new generation
OpennessCode: Open
Weights: Open · MIT
Code, data and weights are all public and the model card gives MIT for the weights; another page says Apache 2.0, so the conflicting code terms are left empty
Benchmarks reported0 counted · none of the 44
Wet-lab validationNone recorded
Unverified fields† claimed_advantage, open_status.code_license, release_milestones.code_release, release_milestones.weights_release
As of2026-11-16

What it does

Uses a modern architecture, cleaned and debiased training data and an efficient codebase to train protein language models far smaller than their contemporaries, to test whether bigger is always better. It is a research baseline rather than a product, positioned against the scale-first route of ESM3 and ESM C.

Benchmarks reported

None of the 44 benchmarks. This model is on the zero-pairing list.

How it evaluates itself: Held-out UniProt pseudo-perplexity; Masked-residue sequence recovery; Unsupervised contact prediction; Disorder discrimination; Inference throughput and design pipeline

Other evaluations (outside the 44)

  • Held-out UniProt pseudo-perplexity (Comparison against baselines only) Self-reported: Held-out likelihood against existing protein language models at several scales, claiming the small model is competitive with the largest baseline
  • Masked-residue sequence recovery (Own task set) Self-reported: Recovery on the human proteome grouped by conservation plus a long-context extension, including the authors' own retrained controls
  • Unsupervised contact prediction (Public benchmark not among the 44) Self-reported: Attention-map contact prediction with logistic regression and random forest: the baseline is better with logistic regression and the gap shrinks with random forest
  • Disorder discrimination (Public benchmark not among the 44) Self-reported: Discrimination of a disorder database against hypothetical proteins, with embedding similarity compared to structure-prediction confidence
  • Inference throughput and design pipeline (Comparison against baselines only) Self-reported: Proteins per second on one GPU, claimed more than an order of magnitude faster than the largest baseline, plus sequence recovery in a folding-and-inverse-folding pipeline

Developers report: Developers report that a 350M model matches or arguably beats a baseline with 43x more parameters on pseudo-perplexity, with about 17x less training compute and much faster inference, attributing this to data quality.

Independent evidence: No independent evidence found. A row-by-row check of the official board found no entry, and no third-party evaluation was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesThree public databases combined into one public dataset: a sequence database, an antibody database and a structural-classification database source
b. Private or licensed dataNoNo private or licensed data reported
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An academic-industry collaboration with no separate company: an AI research institute in Montreal, Canada and a lab at a local engineering university, together with the research arm of a large US biopharma company headquartered in California.
  • Funding: The paper does not itemise funding; nothing inferred
  • Products and partnerships: Releases 120M and 350M sizes together with the training dataset and is included in a GPU vendor's bio-model framework. Its research question targets the scale-first route represented by ESM3 and ESM C.
  • Related models: ESM3, ESM C, SaProt

Reading notes

  1. Its contribution is methodological; it has no entry on the official zero-shot board and cannot enter board comparisons.
  2. The "beats the largest baseline" claim is limited to "some tasks" with no public task list or figures.
  3. Most self-evaluations are own-built or off-site benchmarks, not comparable with the 44 on this site.
  4. Fully open, dataset included, so it is cheap to reproduce and useful as a control baseline.

Caveats

  1. Zero pairings: under this round's relation rules it has no established pairing with the 44 site benchmarks.
  2. The task list behind the claimed advantage is not public and cannot be checked (unverified).
  3. Licence statements conflict: MIT for the weights, Apache 2.0 elsewhere.
  4. No wet lab and no third-party evaluation.

Sources

  1. AMPLIFY preprint (v1 2024-09, v2 2026-01) (preprint)
  2. Weights and model card (120M) (repository)
  3. Public training dataset (repository)
  4. Official repository commit history (repository)
  5. Institute explainer page with the developer framing (official site)
  6. Third-party framework documentation (official site)
  7. Preprint version metadata (archive)
  8. ProteinGym official summary table, checked row by row with no entry (leaderboard)

Corrections

2 corrections
FieldWasNowBasisFound
latest_update.event_kindRecorded as a new versionChanged to a paper revision: 2026-01-05 is preprint v2, not a new model version, and no journal version existssource2026-11-16
latest_update.descriptionSaid only "v2 revision" plus one code changeAdds the v1 and v2 dates, states there is no journal version and that code changes are minor with the last commit on 2025-07-17source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.