Overview
| Name | TxGemma · aliases: TxGemma, txgemma-9b-predict |
|---|---|
| Full name | The name is a product name; the model card gives no expanded form |
| Ambiguous names (not auto-linked) | Gemma |
| Family | Part of the Gemma family |
| Organization | Google DeepMind, Google Research |
| Organization type | Big-tech lab |
| Country | United Kingdom, United States |
| Model class | Drug-discovery LLM (instruction-tuned) |
| Domain | Small molecules (also: Text, LLMs & agents) |
| First release | Basis: Developer blog announcing the open weights Note: The model hub repository creation date of 2025-03-21 is not used as first release under this round's rule; the paper v1 is 2025-04-08 source |
| Other milestones | |
| Latest update | · maintenance only Basis: Last modification date of the three model hub repositories That date is a repository modification, not a new model version; no later TxGemma version was found as of 2026-11-16 source The most recent sourced event, not a claim that nothing changed since. |
| Status | Maintained only · No new version within 12 months, only repository modifications |
| Openness | Code: Partly open Weights: Research use only Weights are open subject to the developer's health-AI terms; the released version was trained only on commercially licensed data, unlike the paper version |
| Benchmarks reported | 1 counted (self 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † open_status.weights_license, open_status.code_license |
| As of | 2026-11-16 |
What it does
Rewrites the tasks of a public therapeutics dataset library as instruction data and fine-tunes an open general base so that one model covers classification, regression and generation across small-molecule and protein therapeutic entities; the predict variant gives structured output, the chat variant natural-language explanation, and the paper adds an agent form. It is the open successor to the same organisation's earlier closed drug LLM and is often used as a comparator in papers such as Enchant and ether0.
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.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 04 TDC (ADMET + DTI) | Developer report, official protocol | — | Self-reported | 66 TDC tasks, recommended splits · evidence |
Other evaluations (outside the 44)
- Agent-form task demonstrations (Own task set) Self-reported: The paper wires the model to tools for multi-step task demonstrations, author-defined and outside the benchmarks on this site
Developers report: Developers report that a single open-weights model covers most tasks of that dataset library, matching or beating specialist models on most of them.
Independent evidence: A company ran it as a comparator on its own kinase task set: a third-party evaluation, though not on an official split.
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | 66 tasks and over 15 million data points from a public therapeutics dataset library source |
| b. Private or licensed data | No | The public weights use only commercially licensed datasets; the paper version also used non-commercial ones source |
| c. Own wet-lab data used for training | No | source |
| d. Wet lab used only for validation | No | source |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The developer is a large technology company's AI research unit, founded in London in 2010, acquired in 2014 and merged with the company's other research team in 2023-04; this model was released through its health-AI developer programme as internal research.
- Funding: Not disclosed
- Products and partnerships: Its predecessor was the same organisation's closed drug LLM on a different base; the base here is the company's open general model; related structural-biology work is AlphaFold 3.
- Related models: AlphaFold 3, Enchant v1/v2, ether0
Reading notes
- Its selling point is covering many tasks in one model, not topping any single task.
- The paper keeps separate counts against specialist and against general-purpose comparators; the two must not be merged when quoted.
- The paper version and the public weights were trained on different data, so reproducing the paper with the open weights may fall short.
- Training and evaluation come from the same dataset library, so transfer to real projects is unverified.
Caveats
- First release follows the official blog; the model hub's repository creation date is not used as evidence.
- 2025-04-10 is a repository modification date and must not be presented as a new version.
- The third-party kinase task set is not that library's official split, so its figures cannot be listed beside official-protocol ones.
Sources
- Paper preprint (preprint)
- Model card (official page)
- Launch blog (blog) · 2025-03-25
- Model hub repository (model hub)
- Third-party evaluation post (blog)
Corrections
2 corrections
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.