retro-lm playground

retro-lm. Plans syntheses: proposes what reacts to make a molecule, one step or a full route. A 0.6B model, run live next to frontier LLMs.

Connecting…

Benchmark

Published results on 500 USPTO-50K test reactions, one step back.

· ·

Target molecule (SMILES)

Try:

No accuracy for a molecule you type: there is no recorded answer to score it against. Cost and speed are measured live.

Draws one of the 500 benchmark reactions. Each model sees only the product; the gateway checks its answers against the reactants recorded in the patent, so accuracy here is measured, not quoted.

Target molecule (SMILES)

Try:

Budget Engine

Asks retro-lm for one step at a time and searches (Retro*) until every branch ends in a molecule from a 23M-molecule purchasable catalogue, within the call budget and 60 seconds. A found route is a plausible plan, not a checked procedure.

Models:

Target

The molecule to make.

The plan the search is pursuing

calls 0 time 0.0 s queueing

Live results

Each model's top proposal. Answers arrive as each model finishes.

Live results from each model appear here: proposals with structures, latency and cost.

Full table

Model	Top-1	Top-10	$ / 1,000	Median s

500 USPTO-50K test reactions, one step back.

How to read these numbers

Top-1 means the first proposal matches the reactants recorded in the patent; top-10 means a match appears anywhere in the ten. Both sides are canonicalized with RDKit before comparing. A chemically valid alternative counts as a miss, so these are lower bounds.
Live accuracy counts only benchmark reactions you ran in this browser session. With a handful of reactions it is noisy; the published 500-reaction figure is the stable one.
Cost for the LLMs is list price for the tokens each request used. For retro-lm it is GPU time, on two different bases: the published figure was measured on an H200 at an assumed $4 per GPU-hour with requests batched, while the live figure is this deployment's own GPU at its configured hourly rate (both named under the charts), one request at a time. They are different hardware and prices, so compare live with live and published with published.
Speed is wall time for the request, measured by the gateway. A cached answer returns at once, so the speed bar shows the time of the original run and is marked cached.
vLLM and SGLang serve the same retro-lm weights with beam search, so their proposals should largely agree and the interesting difference is speed. The published accuracy is the model's own, from an offline beam-search run; the published cost and speed were measured on vLLM, so SGLang shows a dashed "pending" bar there until it is benchmarked.
The frontier LLMs get the same 5-shot prompt as in the published comparison, at their lowest supported reasoning setting ("low"), because these APIs cannot switch reasoning off.
USPTO-50K is public and years old, so memorisation by any model, retro-lm's base model included, cannot be ruled out.

What this does not tell you

It proposes reactants only. It says nothing about conditions, temperature, solvent, yield, side products or hazards.
The benchmark scores one step back, the reactants for the final reaction. The full-route mode chains such steps with a search, but a found route is only as good as each proposed step; nothing checks that it works in a flask.
Training data is patent chemistry, weighted toward medicinal chemistry. Other reaction classes are untested.
A proposal is a starting point for a chemist, not a validated procedure.