# CellFM

Source: https://onco.cc/technologies/cellfm/  
OnCo record `cellfm` (Technology). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

CellFM is an 800-million-parameter single-cell model trained on 100 million human cells.

## Summary

CellFM is a single-cell foundation model from a Chinese consortium, a transformer with efficient attention over gene tokens that scales to 800 million parameters, among the largest single-cell models by parameter count. The 2024 bioRxiv preprint trained it on 100 million human cells and reported gains on cell-type annotation and perturbation prediction tasks. It is a research model for computational biologists exploring whether scale alone improves single-cell prediction. Independent evaluation is limited, and the wider field has found that many single-cell foundation models barely beat simple baselines on perturbation tasks, so its reported gains need external replication. For a newcomer: CellFM is one of the biggest models of its kind, but outsiders have not yet confirmed how much the extra size helps.

## Fields

- Kind: Technology
- Status: emerging
- Last checked: 2026-09-08
- Tags: foundation-model; virtual-cell
- Principle: CellFM is a transformer with efficient attention over gene tokens.
- Since: 2024
- Strengths: Scale
- Limitations: Limited independent evaluation

## Sources

- bioRxiv 2024: https://www.biorxiv.org/content/10.1101/2024.06.04.597369v1

## Connected records

- technologies: [scFoundation (BioMap)](https://onco.cc/technologies/scfoundation/)
- fronts: [AI & Computation](https://onco.cc/fronts/ai-computation/)

---
JSON: https://onco.cc/api/v1/entities/cellfm.json