AI for Bio · Model

xTrimoPGLM

A 100B unified protein language model from BioMap and partners for understanding and generation, with a folding branch

ProteinsMixedChina, United Arab EmiratesMaintained only中文

Overview

NamexTrimoPGLM · aliases: xTrimoPGLM, xTrimoPGLM-100B, xT-Fold, xTrimoPGLM-Fold
Full nameThe paper gives no expanded name
Ambiguous names (not auto-linked)xTrimo
FamilyNot a family entry. Several sizes and the folding branch are treated as versions within one record
OrganizationBioMap, Tsinghua University, MBZUAI
Organization typeMixed
CountryChina, United Arab Emirates
Model classProtein LM / folding
DomainProteins (also: Structure & co-folding)
First release · paper revision
Basis: bioRxiv preprint v1 source
Other milestones
  • First public evidence: · bioRxiv preprint v1 source
  • Preprint v1: · bioRxiv preprint v1 (arXiv version 2024-01) source
  • Journal online: · Nature Methods online source
  • Code release: † · First GitHub commit (reference only, not used as first release) source
Latest update · journal
Basis: Nature Methods online
Journal version published; arXiv v2 had been posted on 2024-12-09 source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Most recent sourced event is the 2025-04 journal paper
OpennessCode: Partly open
Weights: Unclear
Some sizes have weights on Hugging Face and several are in the ProteinGym table; licence not retrieved this round
Benchmarks reported4 counted (board 1, self 1, retrospective 2)
Wet-lab validationNone recorded
Unverified fields† open_status.weights, open_status.code_license, open_status.weights_license, release_milestones.code_release
As of2026-11-17

What it does

Combines masked-language and autoregressive (GLM) objectives in one 100B-parameter model for both protein representation and sequence generation; the xT-Fold branch predicts structure from language-model representations. Compared with ESM3 it takes the single very large model route, far larger than contemporaneous ESM models.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 4 of 4 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
02 ProteinGymOfficial leaderboard or results—BoardEntry on the official summary table · evidence
26 FLIP / FLIP2Developer report, official protocol—Self-reportedSelf-reported after fine-tuning on the random control split · evidence
01 CASPRetrospective dataset use—Third-partyPost hoc on CASP15 (n=56) · evidence
15 CAMEORetrospective dataset use—Third-partyCAMEO 2022-04 to 06 static set, n=194. · evidence

Other evaluations (outside the 44)

  • 18 protein understanding tasks (Own task set) Self-reported: The authors report beating prior state of the art on 15 of 18; mostly fine-tuned, not comparable with zero-shot boards
  • Folding branch vs single-sequence methods (Comparison against baselines only) Self-reported: Compared with ESMFold and OmegaFold; still behind MSA/template methods (author-reported)

Developers report: one 100B model covers understanding and generation, beats prior state of the art on most understanding tasks, and its folding branch beats other single-sequence methods.

Independent evidence: The ProteinGym maintainers' summary table lists several sizes and shows the 100B size no better than smaller ones, consistent with the observation that pLM scaling plateaus after roughly 1B to 4B.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesAbout 939M protein sequences, 200B tokens: UniRef90 plus ColabFoldDB clusters and members source
b. Private or licensed dataNoThe paper lists only two public databases source
c. Own wet-lab data used for trainingNo source
d. Wet lab used only for validationNo source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: BioMap was founded in Beijing in 2020 as an AI life-science company that positions itself as an enabler rather than a drug developer; paper authors also come from Tsinghua University and MBZUAI.
  • Funding: A $100M Series A in 2021 led by GGV; a 2024-06 strategic round of undisclosed size; a 2026-03 media report of a confidential Hong Kong IPO filing, not announced by the company.
  • Products and partnerships: A 2023-10 Sanofi deal for AI modules ($10M upfront, milestones over $1B); the series also includes xTrimoSCPerturb/scFoundation, and a multimodal xTrimo V3 was announced in 2026-10.
  • Related models: ESM3, xTrimoSCPerturb / scFoundation

Reading notes

  1. The 100B scale shows no edge on ProteinGym; in the official table it trails the group's own 3B size.
  2. Folding results are small-sample, self-reported post-hoc evaluations, not CASP/CAMEO entries.
  3. FLIP covers only the GB1 subtask after fine-tuning and does not mean leading on the full suite.

Caveats

  1. "Ahead on 15 of 18" is author-reported and mostly fine-tuned.
  2. CASP15 and CAMEO are post-hoc data reuse, not counted as participation.
  3. Weight availability and licences were not checked item by item.
  4. An earlier site figure of "about 94B unique sequences" was an order-of-magnitude error and is corrected from the paper.

Sources

  1. arXiv preprint (preprint)
  2. Nature Methods paper (journal)
  3. bioRxiv version record (preprint)
  4. ProteinGym official summary table (leaderboard)
  5. Sanofi collaboration press release (press)

Corrections

2 corrections
FieldWasNowBasisFound
first_release2024-01-11(arXiv v1)2023-07-06(bioRxiv v1)source2026-11-16
training_data.a_publicAbout 94B unique sequencesAbout 939 million sequences, 200 billion tokenssource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.