Model release · Canis M
Canis M is a molecular model trained on quantum tokens. A quantum token is a quantum-mechanical representation of a molecule, minted once by our engine and read by the model in place of atomic coordinates. It reaches chemical accuracy on 99.6 percent of held-out molecules at 320× the label efficiency of the standard descriptor.
Ask a chemist what a molecule will do and the honest answer is that it depends what you can afford. The energy that decides whether a drug candidate binds, whether a battery material is stable, whether a reaction runs at all, is obtainable in principle from quantum mechanics and expensive in practice to obtain at all. A single accurate number can cost hours of supercomputer time. A model that learns from those numbers needs thousands of them.
That arithmetic, rather than any shortage of ideas about architecture, is what has governed molecular machine learning for a decade. Ten thousand labelled molecules is a cluster and a week. A second property is another cluster and another week. Teams need tens of thousands and can afford hundreds, so the field has done the sensible thing and built larger models to squeeze more out of the labels it has.
We went the other way and changed what the model reads.
Almost every molecular model begins from the same object: a list of where the atoms sit. Their coordinates, their elements, nothing else. From that, a network is expected to work out the geometry, then the electronic structure the geometry implies, then the energy that follows from the electronic structure. Three inferential steps, and the network pays for all three in labels, because the only way it learns them is by seeing enough examples to reconstruct the physics on its own.
The electrons are the part that matters. Where the atoms sit is a proxy for what the electrons are doing, and what the electrons are doing is what determines the energy. Handing a model coordinates is handing it a photograph of a building and asking it to infer the wiring.
A quantum token hands it the wiring. It is a compact vector carrying a molecule’s correlated electronic structure, minted once by our engine on ordinary GPUs, with no quantum hardware anywhere in the loop.
The cheap calculation still runs. What changes is what the model is asked to do with it: rather than predicting an energy from scratch, it predicts the residual, the part the affordable method cannot reach. The tractable physics stays where it is already cheap, and the model supplies only the difference.
The effect is easiest to see by holding everything else still. Same learner, same scaffold-disjoint split, same budget, three seeds. Only the features change.
A model trained on 50 molecules with tokens reached 0.910 mHa of held-out error. The same model trained on 16,000 with the standard descriptor reached 1.244. That is a 320-fold difference in labelled data, and the smaller run is the more accurate one. It held on every shard we tried.
That is the label efficiency: the same accuracy for 320 times less labelled data. And the crossing is not the interesting part. The slopes are. The classical descriptor improves from 2.29 to 1.24 as its training set grows 320 times, and then flattens; the token curve is still falling at the right-hand edge. The gap is a difference in rate rather than a fixed offset, which is why buying more data does not close it.
Mean error is a comfortable number and a misleading one: a model can post a respectable average while being badly wrong on a minority of molecules, and in a screening pipeline that minority is exactly what reorders your ranking. So it is worth looking at every molecule rather than their mean.
At the 1.6 mHa line that chemists call chemical accuracy, 99.6 percent of held-out molecules fall inside with tokens, against 81.1 percent with the classical descriptor. Put the other way round, the share of molecules you cannot trust drops from roughly one in five to one in 250. The cheap calculation on its own manages about 4 percent. Canis M does this with 162,243 parameters in 645 kilobytes, a three-layer network small enough to email that runs on a CPU without a GPU anywhere near it.
Cheap quantum-chemical methods are not uniformly wrong. They are close enough to trust on easy molecules and badly wrong on hard ones, and the difficulty is not visible in the output: the bad answers look exactly like the good ones, and they cost the same to produce.
Every token carries a difficulty score, a by-product of minting it rather than a second calculation. Sorted by that score into five equal groups, the cheap method’s own error runs from 0.24 mHa on the easiest fifth to 9.76 on the hardest, six times outside chemical accuracy. The classical descriptor tracks that difficulty and loses the threshold in the hardest group. Canis M holds it across all five.
Which turns a flat spend into a directed one. The expensive calculation goes to the molecules that will change the ranking, and not to the ones that were never in doubt.
None of this requires abandoning the representation a team already uses. Concatenate the token to an existing descriptor and train exactly as before: on ANI-2x’s own features, at the same training-set size, accuracy improves 4.4 to 5.3-fold.
X = np.hstack([your_descriptor, siriusq_tokens])
model.fit(X, y) # nothing else changesEverything above is measured between 10 and 14 heavy atoms. Drug candidates are larger, and methods that look good on small benchmarks have a habit of thinning out on real compounds, so the same experiment was run where it counts: 40,600 molecules at 20 to 52 heavy atoms, three seeds, scaffold-disjoint, with only the input features changing between arms.
| Input features | Within chemical accuracy |
|---|---|
| Cheap method alone | 4.4% |
| Composition only | 36.0% |
| Morgan fingerprint + descriptors | 62.6% |
| ANI-2x descriptor | 65.0% |
| Quantum tokens | 93.7% |
Against the strongest classical arm the improvement is 2.97×, and because the arms are paired molecule for molecule it can also be stated directly: the tokens are closer on 78.6 percent of the test set.
The composition-only row is the control that matters. Label size drifts with molecule size, so a model that learned nothing except how big each molecule was would still post a score. At 36.0 percent it does not come close, which rules out the obvious artifact.
The token describes frontier electronic structure, so it carries no information about molecular size and loses to simply counting atoms on size-dependent properties. It is minted per geometry rather than per molecule, so molecular dynamics is not a target. The two size ranges above are separate fits on different windows, and nothing between 14 and 20 heavy atoms is measured. All of it, with the controls that diagnose it, is in the paper.
This is for teams training their own models on expensive physical data: teams developing drugs with AI, frontier model builders in life sciences, molecular biotech, microbiology labs, materials development. If the binding constraint is what a label costs rather than how large the network is, this is an input that can be tested against an existing baseline without changing anything else.
The method is set out in full in Tokenising quantum data for label-efficient training of molecular models. Canis M runs on Hugging Face, and the datasets are public and downloadable today.