Is ChefNMR2 state of the art?Pending benchmarks.Announced Oct 31, 2026; no independent results yet.
A generative model from Ellen Zhong's lab at Princeton, with Princeton Chemistry and Edison Scientific, that proposes the structure of a small molecule from its molecular formula and its 1D 1H and 13C NMR peak lists. It is an atomic flow-matching model: it turns noise into 3D atom coordinates, bonds are assigned afterwards, and repeated sampling gives each candidate structure a confidence. About 288M parameters per checkpoint; MIT inference code, CDLA-Sharing-1.0 weights on Hugging Face and a free web app. ChefNMR2-Agent, which pairs it with LLMs, is offered in Edison Scientific's Kosmos. Not an LLM.
The details so far
- Lab
- Zhong Lab (Princeton University)
- Announced
- Oct 31, 2026
- API
- Not yet · None; it runs locally or in the Hugging Face web app
- Open weights
- Published · ziyux123/chefnmr2 on Hugging Face (the first author's account), not gated: 10 checkpoints from epochs 949 to 1399, each an EMA-folded fp32 model.pt of 1.15 GB (11.5 GB in all)
- Parameters
- About 288M per checkpoint (287,981,888); the paper's setup samples from 10 checkpoints of one training run
- Task
- Small-molecule structure elucidation: molecular formula (with hydrogens) plus 1D 1H and 13C NMR peak lists in; candidate structures as SMILES, ordered by how often the samples give them, out. Planar structure only; stereochemistry is not predicted
- Architecture
- Atomic flow-matching model: an NMR peak encoder (4-layer transformer) and a 12-block multimodal diffusion transformer (hidden size 768, 8 heads) with joint attention between peak and atom tokens; it generates atom coordinates, then bonds are assigned. Covers 14 elements (C, H, O, N, S, P, F, Cl, Br, I, B, Si, As, Se)
- Confidence
- The share of samples that give each distinct planar structure; the paper's adaptive setting draws 100 samples, then 1,000 or 3,000 when they disagree
- License
- Code MIT (ml-struct-bio/chefnmr2); checkpoints CDLA-Sharing-1.0. The training data include NPAtlas and NP-MRD structures (CC BY-NC 4.0), and the lab says to check those terms before commercial use. The preprint is CC BY-NC-ND 4.0
- Code
- Inference code and the paper's five test sets in ml-struct-bio/chefnmr2 on GitHub (Python 3.11, PyTorch 2.5.1); training code and training data are to follow with the journal publication
- Web app
- Hugging Face Space ziyux123/chefnmr2 (Gradio on ZeroGPU), also reachable as an MCP tool from AI chat apps
- ChefNMR2-Agent
- A system built on ChefNMR2: LLMs that reason over its candidates with chemical-shift prediction, spectral interpretation and literature search. Offered in Edison Scientific's Kosmos platform; the open release is listed as coming soon
- Developers
- Ellen Zhong's lab (Princeton Computer Science), with Mohammad Seyedsayamdost's group (Princeton Chemistry) and Edison Scientific; first author Ziyu Xiong
- Earlier version
- ChefNMR1, an atomic diffusion model (arXiv 2512.03127, Dec 2025; NeurIPS 2025), not tracked here
- Lab-stated limits
- Needs the molecular formula; does not predict stereochemistry; accuracy falls for larger molecules and for molecules less similar to the training data
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: across five test sets totaling 2,031 molecules, ChefNMR2 puts the reported planar structure first for 82.3% of molecules, against 35.4% for NMRTransformer, 60.1% for the retrieval method NMR-Solver and 52.0% for Claude-Opus-5, the highest of the six LLMs tested.
Lab claim: it was trained on 2.52 million molecules with experimental and simulated spectra; its sampling-frequency confidence has an expected calibration error of 1.3 percentage points, and it has up to 2,170 times lower median runtime and 360 times lower median cost per molecule than the LLMs tested.
Lab claim: ChefNMR2-Agent elucidated four previously unreported natural products (nocobactin AI-1 and AI-2, hybridamides A and B), matching the structures determined by chemists.
Sources
PAPERSmall molecule structure elucidation from NMR spectra with generative and reasoning models (bioRxiv)biorxiv.org/content/10.64898/2026.10.29.762711v1 · Oct 31, 2026LABChefNMR2 project page (Princeton)chefnmr.cs.princeton.edu/LABEllen Zhong's announcement thread on Xx.com/ZhongingAlong/status/2117354979552043112 · Nov 2, 2026LABSam Rodriques (Edison Scientific) post on X about ChefNMR2 in Kosmosx.com/SGRodriques/status/2117359453330956289 · Nov 2, 2026REPOml-struct-bio/chefnmr2 on GitHubgithub.com/ml-struct-bio/chefnmr2REPOziyux123/chefnmr2 weights on Hugging Facehuggingface.co/ziyux123/chefnmr2REPOChefNMR2 web app (Hugging Face Space)huggingface.co/spaces/ziyux123/chefnmr2Tracking continues until an independent board lists it.