AI for Bio · Model

Intern-S

Open multimodal scientific foundation series layering scientific corpora and specialised tokenisers on general LLMs, including a trillion-parameter mixture-of-experts version

Text, LLMs & agentsNonprofit instituteChinaActive中文

Overview

NameIntern-S · aliases: Intern-S1, Intern-S1-Pro, Intern-S2
Full nameA series name with no unified expanded form given officially
Ambiguous names (not auto-linked)Intern
FamilyPart of the Intern-S family
OrganizationShanghai AI Laboratory
Organization typeNonprofit institute
CountryChina
Model classGeneral scientific multimodal LLM
DomainText, LLMs & agents (also: Small molecules, Proteins)
First release† partly verified · preprint
Basis: Paper v1 of the first generation; a repository creation date cannot serve as first release under this round's rule
Note: The site recorded 2025-07-24, a model hub repository creation date; a 2025-07 public launch has no verifiable page, so the paper v1 is used source
Other milestones
  • First public evidence: † · First-generation paper v1 source
  • Preprint v1: † · First-generation paper v1 source
  • Code release: † · Official code repository source
  • Weights release: · First publication of the second generation's release weights, with a quantised variant two days earlier source
Latest update · new weights
Basis: First publication of the second generation's release weights
That date is the first publication of the weights; the repository's 2026-09-24 last-modified timestamp is only a card edit, not a weights event. Whether the same organisation's separate 2026-10 and 2026-11 model line belongs to this series could not be confirmed source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · A new generation's weights were released within 12 months
OpennessCode: Open · Apache-2.0
Weights: Open · Apache-2.0
Weights of each generation are public on the model hub with quantised variants and a domestic mirror; the code repository carries the same permissive licence
Benchmarks reported0 counted (unconfirmed 6)
Wet-lab validationNone recorded
Unverified fields† first_release, release_milestones.paper_v1, release_milestones.code_release, training_data.b_private_licensed
As of2026-11-16

What it does

Continues pretraining a general large model on scientific corpora with specialised tokenisers and a vision encoder, then post-trains with multi-task and agent reinforcement learning across molecules, reactions, materials and microscopy images; later versions scale the expert count to a trillion parameters and add a domain-specialisation module that leaves the backbone untouched. It divides work with the same organisation's SciReasoner, a smaller model dedicated to structured reasoning, and unlike biology-specific bases such as Evo 2 it does not pretrain on sequences from scratch.

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. 0 of 6 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
14 PubMedQAUnconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · evidence
18 LAB-Bench / LABBench2Unconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · evidence
23 BioASQUnconfirmed (not counted)—Unconfirmed sourceNo report by it was found in the official material. †disputed · evidence
29 BixBenchUnconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · evidence
41 BioML-benchUnconfirmed (not counted)—Unconfirmed sourceNo report on this benchmark found in official materials. †disputed · evidence
42 BioMysteryBenchUnconfirmed (not counted)—Unconfirmed sourceNo report on this benchmark found in official materials. †disputed · evidence

Other evaluations (outside the 44)

  • Instruction-style sequence and molecule benchmarks (Public benchmark not among the 44) Self-reported: Tasks on two public instruction datasets compared with several frontier general models, all self-reported
  • Protein QA, microscopy and chemistry benchmarks (Public benchmark not among the 44) Self-reported: Several evaluations outside this site's 44 with mixed outcomes, some still topped by closed frontier models
  • An author-defined protein binder task set (Own task set) Self-reported: A task set defined in the second generation's preview paper, compared with frontier general models

Developers report: Developers report open multimodal scientific models that beat frontier general models on scientific sequence instruction tasks.

Independent evidence: No independent evidence found. No outside-lab reproduction was found, and the evaluations use the same organisation's toolchain

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesScientific documents used as page images for visual pretraining, interleaved image-text data and an image retrieval corpus; agent training uses public executable task sets whose task and environment counts the paper lists source
b. Private or licensed dataNot disclosed source
c. Own wet-lab data used for trainingNo source
d. Wet lab used only for validationNoNeither the three papers nor the official repository reports wet-lab validation source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is a government-backed new-type research institution in Shanghai, reported as founded in July 2020; the series began with a vision model built with a company in 2021 and later extended to language and science. Scale and budget are not public.
  • Funding: Government funding; no commercial investment.
  • Products and partnerships: The same organisation also runs SciReasoner and general vision and language model lines; the second generation's preview does reinforcement learning against several open and closed agent harnesses.
  • Related models: SciReasoner, Frontier general-purpose LLMs, Evo 2

Reading notes

  1. Self-reported biology results centre on instruction-style tasks where the general comparators score very low, so the gap mostly reflects task format and dedicated training.
  2. Closed frontier models still beat it on microscopy and molecular knowledge evaluations, so it should not be read as leading across the board.
  3. Evaluations use the same organisation's toolchain, which limits third-party reproduction.
  4. It has no countable pairing with the benchmarks on this site and cannot be used for ranking here.

Caveats

  1. First release does not use the model hub repository creation date; a 2025-07 launch has no verifiable page, so this record uses the paper v1 and marks it unverified.
  2. 2026-09-24 is only a repository last-modified timestamp and must not be presented as a weights event.
  3. Parameter counts differ between the repository and the papers, and sizes of different versions must not be mixed.
  4. Whether the same organisation's other model line from late 2026 belongs to this series is unconfirmed and it is not merged into this record.

Sources

  1. Official code repository (repository)
  2. First-generation paper (preprint)
  3. Trillion-parameter version paper (preprint)
  4. Second-generation preview paper (preprint)
  5. Second-generation model collection (model hub)

Corrections

2 corrections
FieldWasNowBasisFound
latest_update2026-09-242026-09-13 (weights first public; 09-24 is only the repository's last-modified time)source2026-11-16
first_release2025-07-24 (model-hub repository creation date)2025-08-21 (first-generation paper v1), marked unverifiedsource2026-11-16

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