Announcing Primateria

Today we are announcing Primateria, an AI-native scientific lab.

We're starting with one of the hardest problems in quantum mechanics: the electronic Schrödinger equation. Our first model tackles it in transition metals, a notoriously difficult regime. You can try it now.

Every dot is one electron position sampled from |Ψ|², shown for MnO₂, FeCl₂ and other transition metal systems.

A recursive research loop

Despite the challenge, frontier models were able to help decompose the research problem into executable steps, implement them, and orchestrate experiments across our cluster of 68 RTX Pro 6000s.

Each run produced numerical feedback that became context for the next step, creating a recursive research loop. This computational feedback from GPUs kept the loop grounded in physics.

P2, a foundation model for transition metal chemistry

The loop produced P2, our foundation model trained across transition metal chemistry.

On eight held-out molecules, warm starting from P2 reached 265 mHa lower energies than cold starts after 5–10k optimization steps.

We then tested transfer across bond breaking for CrF, which was never in the training set. After adaptation, we recover a dissociation energy of 4.58 eV against 4.6 eV experimentally, with 3.8 kcal/mol RMSE in the well against NEVPT2.

CrF X⁶Σ⁺ potential energy curve, r ≤ 4.3 Å: blind P2 against a quantum chemistry reference

Well RMSE against NEVPT2 over the 12 radii from 1.53 to 2.08 Å: separate 9.8 mHa (6.2 kcal/mol) · joint 8.9 mHa (5.6 kcal/mol). Dissociation energy D_e from a Morse fit to the spin-pure points: 4.58 eV, experiment ≈ 4.6 eV.

Data table

r (Å)	Well	NEVPT2 ΔE (mHa)	P2 separate ΔE ± SE (mHa)	P2 joint ΔE ± SE (mHa)	⟨S²⟩ separate	⟨S²⟩ joint
1.42		+55.9	+78.4 ± 1.7	+89.4 ± 1.8	8.760	8.766
1.53	yes	−13.5	−1.9 ± 1.7	+7.2 ± 0.6	8.761	8.767
1.58	yes	−30.5	−19.6 ± 0.6	−15.0 ± 1.2	8.760	8.767
1.63	yes	−41.8	−31.5 ± 1.2	−30.2 ± 0.9	8.759	8.768
1.68	yes	−48.5	−43.4 ± 0.5	−40.8 ± 0.9	8.763	8.766
1.73	yes	−51.7	−37.4 ± 1.0	−47.0 ± 1.4	8.773	8.766
1.78	yes	−52.7	−38.4 ± 0.9	−50.3 ± 1.8	8.783	8.766
1.83	yes	−52.1	−47.2 ± 1.3	−51.7 ± 0.9	8.760	8.765
1.88	yes	−49.6	−46.3 ± 1.9	−49.5 ± 0.3	8.765	8.767
1.93	yes	−45.9	−44.4 ± 1.5	−48.2 ± 0.4	8.768	8.766
1.98	yes	−41.1	−34.1 ± 0.5	−44.8 ± 0.6	8.774	8.766
2.03	yes	−35.6	−19.1 ± 0.3	−39.4 ± 1.2	8.765	8.766
2.08	yes	−29.7	−30.4 ± 0.8	−34.2 ± 1.1	8.772	8.765
2.30		−1.3	−7.6 ± 0.8	−10.2 ± 0.6	8.761	8.764
2.55		+42.7	+12.3 ± 0.7	+11.8 ± 0.9	8.763	8.762
2.85		+79.8	+39.3 ± 1.6	+42.4 ± 1.1	8.760	8.760
3.20		+98.1	+65.3 ± 1.5	+70.7 ± 0.7	8.756	8.760
3.70		+107.1	+93.5 ± 0.7	+100.9 ± 1.7	8.756	8.759
4.30		+110.4	+112.4 ± 0.9	+138.8 ± 1.3	8.756	8.777

CrF X⁶Σ⁺ potential energy curve: P2, which never saw a Cr–F bond in training, against the NEVPT2 reference computed with the same Hamiltonian. Every curve is shifted by its own mean over the radii shown, so the comparison is of shape.

We think this method generalizes

Many computational science problems can be dissected into subproblems that are executable and verifiable by physical truth. As models improve, more of these sequences can be searched and executed autonomously.

Over time, we are building AI systems to take on increasingly large parts of that loop, while keeping each step tied to reproducible computation. Quantum chemistry is just our first arena.

What's next

These results are early. We are pushing into Cr₂, larger valence spaces, property extraction, and periodic systems next, before tackling adjacent scientific problems. We will release the methods and full benchmarks in a preprint soon.

For now, try the model and send us systems you think should break it. If you find this interesting and you'd like to help, join us.