DP Technology

The directions that can currently be confirmed are: a peer-reviewed publication record no other private company tracked here approaches, a multi-backend open-source simulation stack governed by a separate community, a set of models whose strongest results are tied to experimental measurements rather than to predictions in the abstract, a broad commercial product line with named enterprise customers and no published outcome, and a $114 million Series C. A Europe PMC search on September 17, 2026 returns 113 indexed papers carrying a DP Technology affiliation, eight of them in Nature.

Scope note: this page tracks the public record only: the company's own pages, the peer-reviewed literature, and third-party reporting. No inside information. It records how far public material goes, not how well the company is doing.

Overview

1. The publication record, and what a count can and cannot say

What can be confirmed: searching Europe PMC for the affiliation "DP Technology" on September 17, 2026 returns 113 indexed papers. By year: 2 in 2021, 13 in 2022, 18 in 2023, 26 in 2024, 35 in 2025 and 19 so far in 2026. By venue: 8 in Nature, 9 in Nature Communications, 9 in Journal of Chemical Theory and Computation, 8 in Journal of Chemical Information and Modeling, 5 in JACS, 4 in Nature Structural & Molecular Biology.

This is the largest peer-reviewed record of any private company on this site, and the gap is not close. It is also a number anyone can regenerate: the query is linked below, and it will keep moving.

What a count of papers is evidence of, and what it is not. It shows the company is embedded in working science: its people are on author lists that referees accepted, in venues that publish results rather than announcements. That is real, and it is participation. It does not show that any product works, because a paper count cannot tell you whose contribution drove which result. Most of these papers are collaborations in which DP Technology is one affiliation among several university and institute affiliations, and the count does not distinguish a corresponding author from a fifth co-author who supplied a simulation. PubMed also under-covers materials science and physics, so the true total across all fields is higher than 113, which cuts the same way: it makes the number a better measure of reach and a worse measure of ownership.

Boundary: participation in published science. Confidence: high for the count and the venues (independently reproducible); low for reading it as evidence about the products.

2. What the company owns, and what it does not

What can be confirmed: DeePMD-kit version 3 was published in Journal of Chemical Theory and Computation on May 2, 2025. It is "a significant update featuring a multibackend framework that supports TensorFlow, PyTorch, JAX, and PaddlePaddle backends," designed so that "seamless back-end switching with minimal modifications" lets other machine-learning-potential packages integrate with it.

DeepModeling is a community, not a product line. It began in December 2017 with the deepmd-kit project, now hosts "over 30 open-source projects," runs the OpenLAM initiative for large atomistic models (under which DPA-1 and DPA-2 were released), and reports "tens of thousands of users from more than 20 countries." It has its own governance and its own contributors.

And one attribution is worth getting right. The 2020 ACM Gordon Bell Prize, for "Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning," is routinely attached to this company's story. ACM named a nine-member team, and the announcement lists Weile Jia (UC Berkeley), Han Wang (Institute of Applied Physics and Computational Mathematics, Beijing), Mohan Chen and Denghui Lu (Peking University), Lin Lin (UC Berkeley and Lawrence Berkeley National Laboratory), and Roberto Car, Weinan E and Linfeng Zhang (Princeton University). No company affiliation appears on it. Co-founder Linfeng Zhang is listed at Princeton.

That prize is real, it is the intellectual lineage this company was built on, and it is not a company result. Keeping those three things separate (what the company owns, what the community owns, what the universities won) is the single most useful discipline for reading DP Technology's record, and the reason this section exists.

Boundary: software and lineage. Confidence: high (peer-reviewed software paper, public community governance, a prize record that names its own affiliations).

3. The models: strongest where they touch a measurement

What can be confirmed: several models have peer-reviewed papers behind them.

• Uni-Mol, a 3D-conformation-based molecular representation framework, in Nature Communications (August 2024).
• XtalNet, "End-to-End Crystal Structure Prediction from Powder X-Ray Diffraction," in Advanced Science (January 2025): predicting structure from an experimental diffraction pattern rather than from composition alone.
• NMRexp, "a database of 3.3 million experimental NMR spectra," in Scientific Data (December 2025), and NMR-Solver, "automated structure elucidation via large-scale spectral matching and physics-guided" refinement, in Nature Communications (April 2026).
• DPA-3 applied to biomolecular force-field optimization, in Journal of Chemical Information and Modeling (May 2026).

The pattern is worth naming. The results that made it into the strongest venues are the ones anchored to an experimental measurement (a diffraction pattern, an NMR spectrum), where the model's output can be checked against something an instrument produced. That is a different and harder kind of claim than a benchmark score, and it is where this portfolio is most defensible.

Boundary: interpretation and prediction. A structure solved from a spectrum is a measurement made usable, not a material or a drug. Confidence: high (peer-reviewed, in venues that publish data).

4. The commercial stack, and what it does not disclose

What can be confirmed: the company describes a "Particle Universe" engine spanning "reading, computing, experimenting," delivered as Science-as-a-Service. The named products are Bohrium (research cloud, with Lebesgue for compute scheduling), Hermite (computational drug design), RiDYMO (hit discovery), Piloteye (battery R&D), Uni-Lab (laboratory integration) and SciMaster (a general scientific agent). The company was founded in 2018 and has offices in Beijing, Shanghai, Shenzhen and Yibin.

Usage figures, stated by the company and reported through coverage of its Series C: tools used by "more than 1,000 universities and research organizations" and "150 corporate clients globally," with PetroChina and BYD named among the largest.

Boundary: a thousand institutions and not one published outcome. No customer KPI, no before-and-after, no named result attributable to any of these products. RiDYMO is described as targeting previously undruggable targets and no candidate, IND or trial is named. Uni-Lab is a laboratory operating system with no published throughput. SciMaster is an agent with no published evaluation. Confidence: medium that the deployment is as described (company-stated, named customers, relayed by third-party reporting); low for any scientific or commercial outcome.

Open leads: any KPI from a named customer; a RiDYMO candidate; an evaluation of SciMaster on a benchmark someone else maintains; Uni-Lab throughput.

5. Capital, and the question nobody has answered

What can be confirmed: a Series C of ¥800 million (about $114 million) announced December 24, 2025, led by Fortune Venture Capital and the Beijing Jingguorui Equity Investment Fund, with the Beijing Artificial Intelligence Industry Investment Fund, the Beijing Pharmaceutical and Health Industry Investment Fund, Lenovo Capital, Incubator Group and Oriza Hua participating. No valuation and no revenue figure were disclosed.

The open question this record poses is about shape, not scale. DP Technology sells a research cloud, a compute scheduler, a drug-design platform, a hit-discovery platform, a battery-R&D platform, a lab operating system and a scientific agent. Either those compound (every customer's work feeding one data flywheel that makes the next model better) or they are several service businesses sharing a brand and a simulation core. Nothing public distinguishes the two, and the products' own disclosures are exactly where that evidence would have to appear.

Boundary: financing. Confidence: high for the round and the investors; the strategic question is unresolved in public.

What could not be found

• No published end-to-end autonomous discovery. Nothing shows SciMaster or Uni-Lab running a loop that produced a finding.
• No customer outcome, KPI or case study for any of the 150 corporate clients.
• No RiDYMO candidate, IND or trial registration.
• No valuation, revenue or headcount.
• No breakdown of the company's contribution across the 113 papers: position in author lists, corresponding authorship, or which used company products.
• No throughput figure for Uni-Lab, and no evaluation of SciMaster on an externally maintained benchmark.

Method and scope

Built from the company's own English about page, the DeepModeling community pages, Europe PMC affiliation searches, the peer-reviewed papers named above, the ACM Gordon Bell Prize record, and third-party coverage of the December 2025 Series C. Figures stated by the company and relayed by reporters are labelled as such.

Confidence is assigned claim by claim, not to a whole direction. The labels:

• High: a peer-reviewed publication, a public artifact, or an independently reproducible query;
• Medium: an official statement or company-reported figure (inspectable in detail, unverified outside);
• Low: a single-source claim, or a figure whose meaning does not survive the use it is put to.

Two operating rules: revisions are dated by version and are not treated as new projects; evidence that a project exists is not evidence that the technology has succeeded.

One remark specific to this company. The 113-paper count is simultaneously the strongest evidence on this page and the easiest to misread. It is strong because anyone can rerun the query and because referees, not marketers, let those papers through. It is easy to misread because a paper count measures participation in science, not outcomes from products, and the distance between "this company's researchers are on Nature author lists" and "this company's platform accelerates its customers' R&D" is exactly the distance a page like this exists to keep visible.

Primary sources

Peer-reviewed

• DeePMD-kit v3: A Multiple-Backend Framework for Machine Learning Potentials, *J. Chem. Theory Comput.* (2025-05-02)
• Data-driven quantum chemical property prediction leveraging 3D conformations with Uni-Mol, *Nature Communications* (2024-08-19)
• End-to-End Crystal Structure Prediction from Powder X-Ray Diffraction, *Advanced Science* (2025-01-04)
• NMRexp: A database of 3.3 million experimental NMR spectra, *Scientific Data* (2025-12-18)
• NMR-Solver: automated structure elucidation via large-scale spectral matching, *Nature Communications* (2026-04-02)
• Europe PMC affiliation search: 113 results on 2026-09-17

Company and community

• DP Technology: About
• DeepModeling: About
• 2020 ACM Gordon Bell Prize announcement, with the winning team and their affiliations