AI for Bio · Model

ether0

A 24B-parameter chemistry reasoning model whose answers are molecules

Small moleculesNonprofit instituteUnited StatesActiveSite entity page (Chinese site)中文

Overview

Nameether0 · aliases: ether0
Full nameThe model carries a lowercase code name with no expanded form
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single weights release with no family of versions
OrganizationFutureHouse
Organization typeNonprofit institute
CountryUnited States
Model classChemistry reasoning LLM (reinforcement-learning fine-tune)
DomainSmall molecules (also: Text, LLMs & agents)
First release · paper revision
Basis: Preprint v1, with weights and the release post alongside
Note: The developer blog is one day later and does not move the first-release date source
Other milestones
  • First public evidence: · Preprint v1 source
  • Preprint v1: · Preprint v1 source
  • Code release: · Last maintenance commit in the repository source
  • Weights release: † · Weights repository made public source
  • Official announcement: † · Developer release post source
Latest update · paper revision
Basis: Preprint v2 revision
The preprint reached v2 on 2 December 2025; earlier events are a repository maintenance commit on 26 October 2025 and acceptance at a conference that year.
Note: The site previously recorded the 2025-10-26 maintenance commit as the latest update source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · A paper revision and maintenance commits within 12 months
OpennessCode: Open
Weights: Open · Apache-2.0
Weights under Apache-2.0, with the self-built evaluation set public as well
Benchmarks reported1 counted (modified protocol 1)
Wet-lab validationNone recorded
Unverified fields† release_milestones.weights_release, release_milestones.official_announcement, open_status.code_license
As of2026-11-16

What it does

Recasts chemistry tasks as verifiable questions whose answer is a molecule, then applies supervised fine-tuning and reinforcement learning to a 24-billion-parameter open general model, spanning solubility, functional groups, reaction prediction, absorption and metabolism, acid dissociation and hundreds of other task types. Unlike instruction-tuned models on public therapeutics datasets such as TxGemma, it follows a reasoning-with-verifiable-rewards route.

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. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
07 USPTO family (+PaRoutes)Developer report, modified protocol—Self-reportedSelf-built reaction prediction, not official test set · evidence

Other evaluations (outside the 44)

  • Self-built chemistry question set (Own task set) Self-reported: A public evaluation set of 25 questions per task type, released alongside the model
  • Multiple-choice comparison with a frontier general model (Comparison against baselines only) Self-reported: Reported side by side with a closed reasoning model on zero-shot multiple choice

Developers report: a specialist chemistry model trained with verifiable rewards that beats a classical reaction model on a self-built held-out task using far fewer samples, with weights and question set fully open.

Independent evidence: No independent evidence found. No third-party evaluation was found, though open weights and an open question set make one possible

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesSeveral public chemistry and pharmacology measurement archives, with questions derived from existing measurements; the paper reports 375 task types and 640,730 questions source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNoNo new experiments; questions come from existing measurements source
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is a non-profit research institute founded in 2023 in San Francisco, funded mainly by one technology-sector donor, producing research agents and open models; a for-profit spinout followed in November 2025.
  • Funding: The non-profit runs on donations; the spinout announced a $70M seed round in December 2025 at a reported valuation of about $250M, co-led by two venture firms.
  • Products and partnerships: The same institute produces literature and experiment agents; related records here include Robin and Kosmos.
  • Related models: TxGemma, Robin, Kosmos / Edison Analysis

Reading notes

  1. The reaction comparison differs in both training size and task construction and cannot be read as a standing on that dataset's official board.
  2. The question set is defined by the developer, so the conclusions cover only chemistry tasks inside that training distribution.
  3. Zero-shot multiple choice sits slightly below a closed reasoning model, showing the gap that remains between a specialist small model and frontier general ones.
  4. Open weights and an open question set are its main transparency advantage, since third parties can recheck directly.

Caveats

  1. Its only pairing here uses a self-built task, not the official test split.
  2. It reports no score on the common small-molecule property boards.
  3. Questions derive from existing measurements; there is no new experimental data.
  4. The latest update is a paper revision; the weights have not been re-released.

Sources

  1. Preprint 2025 (preprint)
  2. Weights repository (repository)
  3. Developer release post (company blog)
  4. Repository commit feed (repository)

Corrections

1 correction
FieldWasNowBasisFound
latest_update2025-10-26 (repository maintenance commit)2025-12-02 (preprint v2 revision)source2026-11-16

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