Literature · futurehouse
ether0
A 24-billion parameter language model that reasons in English to design and modify molecular structures as SMILES.
What ether0 is
ether0 is a 24-billion parameter chemistry reasoning model developed by FutureHouse. Built on Mistral-Small-24B-Instruct-2501, it is specialised to take prompts written in English — which can include chemical structures formatted as SMILES (Simplified Molecular Input Line Entry System) strings — and generate molecular structures as answers.
Unlike broad-domain conversational models, ether0 is trained specifically for targeted molecular generation tasks. These include proposing chemical structures matching targeted chemical properties, carrying out single-step retrosynthesis from commercially accessible reagents, predicting reaction products, and elucidating natural product structures from an organism and molecular formula.
The model operates primarily in an inverse design mode. It does not evaluate forward physical measurements on arbitrary molecules, but instead proposes or alters molecular structures to meet specific biological, safety, or physical criteria.
How it works
ether0 was created by initialising from Mistral-Small-24B-Instruct-2501 and training on reasoning traces from DeepSeek-R1 to establish step-by-step reasoning patterns and template adherence. Following this, multiple independent specialist models were trained using Group Relative Policy Optimization (GRPO) — a reinforcement learning method — with verifiable reward signals dedicated to individual chemical tasks.
Correct reasoning traces from these specialists were filtered and aggregated to fine-tune the core model. A subsequent stage of GRPO was conducted across all supported tasks jointly, concluding with targeted safety post-training to refuse prompts involving poisons, explosives, and compounds listed on OPCW Schedules 1 and 2.
Key facts
- Base model
- mistralai/Mistral-Small-24B-Instruct-2501
- Parameter count
- 24B
- Model architecture
- Mistral
- Input modality
- Text (English instructions and SMILES strings)
- Output modality
- Text (SMILES strings)
- Hardware requirement
- 80 GB GPU memory
What you can use it for
- Property-guided molecular modification
- Proposing alterations to an existing SMILES structure to adjust its aqueous solubility (LogS), acid dissociation constant (pKa), or blood-brain barrier permeability while preserving core scaffolds.
- Single-step retrosynthesis and reaction prediction
- Identifying potential single-step reaction precursors from likely commercially available reagents for a target molecule, or forecasting the chemical outcome of a reaction.
- Natural product elucidation
- Generating candidate chemical structures given an empirical molecular formula in Hill notation and the source organism from which the compound was isolated.
- Bioactivity and safety modulation
- Designing or altering molecules to target human cell receptor modes (such as agonism) or to remove specific GHS toxicity classifications and adjust quantitative LD50 values.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
ether0
Strengths and limits
Good at
- Generates molecular structures that adhere to specific multi-property constraints, including target pKa, solubility, and receptor binding phenotypes.
- Processes non-canonical SMILES strings and preserves specified stereochemical configurations in prompts.
- Uses chain-of-thought chemical reasoning reinforced by verifiable task-specific rewards during training.
Watch out for
- Cannot answer direct forward property prediction questions, such as requesting the numerical pKa of an existing chemical structure.
- Exhibits poor textbook chemical knowledge, limited synonym understanding, and weak performance on benchmarks such as ChemBench.
- Frequently misidentifies common chemical names, such as confusing lysine with glutamic acid, requiring input structures to be explicitly provided as SMILES.
- Provides no uplift on tacit laboratory procedures, including compound purification, processing, or physical reaction scale-up.
Inputs and outputs
You provide
Prompt
The model continues from here.
You get back
The model's completion.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use ether0 commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run ether0 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 2× A100, typically in 12s, and a run costs about 560 credits. Visitors without an account get 1,000 credits a month.
Can ether0 calculate the physical properties of a user-supplied molecule?
No. The model is specifically trained to output molecules rather than compute forward properties. It can generate a molecule that matches a target pKa or modify an existing structure to shift its pKa, but it cannot answer queries asking for the direct pKa value of an arbitrary input molecule.
What is the recommended format for entering molecules into the model?
Molecules should be supplied as SMILES strings. While the model has limited support for IUPAC nomenclature and chemical formulas, it frequently errs when common names are provided.
Do input SMILES strings need to be canonicalised?
No. ether0 accepts non-canonical SMILES strings, and stereochemical designations can be included directly in the input representation.
Can the model assist with experimental laboratory protocols?
No. The model provides 1-step retrosynthetic proposals and reaction outcomes, but it does not offer specialised guidance on experimental scale-up, reaction processing, or purification beyond standard language models.
How to cite it
Narayanan, Siddharth M., Braza, James D., Griffiths, Ryan-Rhys, Bou, Albert, Wellawatte, Geemi P., Ramos, Mayk Caldas, Mitchener, Ludovico, Rodriques, Samuel G., and White, Andrew D. (2025). Training a Scientific Reasoning Model for Chemistry. arXiv preprint arXiv:2506.17238.
Written from ether0's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.