Baiyao Tech (百曜科技) releases AURA CellOS, first AI virtual cell based on LLM-JEPA architecture.
The AMW Read
Introduces a novel architectural approach (JEPA) to the AI virtual cell segment, which updates the ongoing debate about whether first-generation models have hit a scaling plateau. The company is a new entrant on the player map with credible prior performance in the Arc Institute challenge.
Baiyao Tech (百曜科技) releases AURA CellOS, first AI virtual cell based on LLM-JEPA architecture.
Baiyao Tech (百曜科技), a Chinese AI-biology platform company, has released AURA CellOS, described as the world’s first AI virtual cell world model built on an LLM-JEPA (Joint Embedding Predictive Architecture) framework. The model is trained on 390.5 million human single-cell transcriptomes, covers over 40 tissue types and 260 cell types, and contains 12 billion parameters. The company claims CellOS achieves state-of-the-art performance on cell-state annotation and perturbation response prediction, with a Pearson_edist score of 0.619—66% ahead of the previous best open-source model, TranscriptFormer.
The release is significant because it marks an architectural departure from the first generation of AI virtual cell models, which primarily used language-model objectives to predict static gene expression patterns. By adopting JEPA—an approach originally proposed by Yann LeCun for learning world models—CellOS aims to internalize causal relationships between cell states, rather than simply memorizing expression profiles. This development updates Segment 06 (Healthcare/Bio) and touches on the ongoing debate within the field about whether scaling law benefits for single-cell models have plateaued (the article cites a June 2026 Nature Methods study showing performance saturation after 22,000 cells). Baiyao Tech’s claim is that the bottleneck is architectural, not data volume—a position that, if validated, could reshape the competitive landscape for AI-driven drug discovery and cellular engineering.
Industry analysts should watch how Baiyao Tech executes on its stated “data-model-experiment” closed loop. Unlike pure-play AI model providers, the company has built wet-lab infrastructure for high-throughput perturbation experiments, positioning itself as a platform in the style of Xaira Therapeutics or Noetik. The broader AIVC (AI Virtual Cell) segment is still nascent but attracting significant capital and talent, including the Arc Institute’s Virtual Cell Challenge. Baiyao Tech’s team was a top performer in that challenge, lending credibility to its architectural claims. However, the company faces challenges common to the segment: scarcity of high-quality perturbation data, the need for independent third-party validation, and the long road to integrating virtual cell predictions into actual drug development pipelines.