An ex-Princeton postdoc raised $7M for a scientific foundation model
Two gaps got money and code this week: science still has no compiler, and the next generation of AI chips still has no common language.
Keli Xinxu, an AI-for-science startup founded by an assistant professor at Hong Kong University of Science and Technology, has raised nearly 50 million yuan — about $7 million — in a first round led by Inno Angel Fund, with Yijing Capital, Xiaomiao Langcheng and Linge Ventures joining and Lighthouse Capital as strategic incubator and sole financial adviser. Founder and CEO Zhang Zaixi, born in 1998 and trained in the University of Science and Technology of China's junior class, built the molecular screening model MGSSL, the drug-molecule generator FLAG, the protein-pocket designer PocketGen and the RNA drug designer RNAGenesis before moving up the stack to agents. His team's STELLA and BioClaw connect scattered models, databases and lab tools; the company says STELLA has served more than 1,000 researchers, that 70% of its international users come from Stanford, Princeton, Harvard and MIT, and that BioClaw's 1,200-member community has a working relationship with the listed synthetic-biology company Ginkgo Bioworks.
The most concrete claim is a wet-lab one. At a national key laboratory that has spent nearly two decades on acute myeloid leukemia, the company says a privately deployed STELLA returned a ranked list of candidate targets in ten minutes, and that the top-ranked target — which had no direct prior literature — was subsequently validated across several cell lines with preliminary positive results. The same literature search, database cross-check and target shortlist would normally take a researcher days to weeks, and the hit was outside the lab's existing research priorities. Treat it as a company-reported result: STELLA was not the party running the cell-culture validation, and the finding is a starting point, not a published drug target.
Zhang's pitch is that a stronger model is not the bottleneck — "science is the next coding," in his framing, and the missing piece is a model that reasons in the scientific space itself rather than converting structures, images and omics matrices through text and code. Keli Xinxu calls that shift going from language tokens to scientific tokens, and its first foundation model targets the central dogma, DNA to RNA to protein to phenotype. Its yardstick is antibody design, where the industry's best efforts get roughly one or two working sequences out of ten; the target is 80–90%. Alongside that, the company is building a sandbox of more than 200 instruments and training vision-language-action models to drive robot arms through exploratory experiments that fixed high-throughput equipment handles badly, and collecting the failed and intermediate runs — yields, reaction speeds, mid-process images — that never make it into papers. It is the commercial version of an argument from the Princeton AI lab where Zhang did his postdoc, which we covered this month — Princeton's Mengdi Wang: LLMs haven't made a real scientific discovery yet.
Tsingmicro and the Beijing Academy of Artificial Intelligence open-sourced Open3D-PIMC at China's national computing conference, presented as the first programming model and software framework written for 3D compute chips — the vertically stacked memory-and-logic designs meant to cut the distance data travels. Under the framework, a developer declares how data is distributed and how a task is split, and the toolchain handles algorithm-to-hardware mapping, while each vendor's private instruction set stays a plugin rather than a fork. That matters because the hardware already shipped without it: BAAI's AI systems lead says no domestic open-source 3D chip compiler project existed, and without one the silicon cannot reach its rated performance. Open3D-PIMC joins FlagOS, the BAAI-led stack built with Tsinghua, Peking University and the Chinese Academy of Sciences that now covers 20 vendors and more than 30 AI chips, and adapted Qwen3.8-2.4T-A95B across nine of them on release day. Tsingmicro says it has more than 5,000 petaflops of domestic reconfigurable compute deployed or under construction, and a working group has promised new 3D-chip software optimizations in the fourth quarter. The hardware race gets the headlines; the toolchain is what decides which chip a developer's code will run on.
What to watch: whether the antibody-design numbers land near 80%, and whether any 3D-chip vendor outside China adopts Open3D-PIMC.
Would you bet on a science model trained on experiments rather than papers, or is the wet lab still the only verifier that counts? Tell us in the comments.
Sources: 智东西 Zhidx — 中科大少年班95后造"AI科学家"，融资近5000万 · NetEase Tech — 科理新序获近5000万首轮融资，英诺科创基金领投，押注AI4S基模×物理AI · Beijing News — 清微智能主导开源全球首个3D算力芯片编程模型 · CNR — 全球首个3D算力芯片编程模型开源发布
