Press
flylabs trains one fixed neural architecture, the complete fruit fly connectome, for real-world tasks, instead of designing a new architecture per task.
About
flylabs is a connectome-based AI lab. It ships one engine, FlyCore, which runs the 139,248-neuron wiring diagram of a fruit fly brain as a sparse recurrent network, and a catalogue of weight sets trained on that engine for specific tasks: mouse and keyboard control, game controllers, actuators, object detection, audio events, motion tracking, speech segmentation and command grounding.
Standard adapters map screens, cameras, microphones, sensors and embeddings onto the brain's sensory neurons and read its descending neurons back into actions. Classical models act as co-processors where the fly brain has no competence, such as words. Weights are public on Hugging Face and free up to 100,000 monthly active devices per platform.
| What we ship | 1 weight set available, 7 in beta, 1 in development, across control, perception and language. |
|---|---|
| How we build them | The connectome is the fixed architecture. We train weights, adapters and readouts per task by imitation and reinforcement, and keep the wiring untouched. |
| How it ships | PyTorch for training and inference, a Rust runtime for devices and actuators, WebGPU in the browser. |
| Where it runs | Laptops, workstations and robot controllers today; phones and the browser for the perception weight sets. |
| Data | FlyWire connectome (Dorkenwald et al. 2024), annotations by Schlegel et al. 2024 (CC BY 4.0). |
Independent research
Jin et al., Whole-Brain Connectomic Graph Model Enables Whole-Body Locomotion Control in Fruit Fly, arXiv 2602.17997, 2026. Lappalainen et al., Connectome-constrained networks predict neural activity across the fly visual system, Nature 2024. Shiu et al., A Drosophila computational brain model reveals sensorimotor processing, Nature 2024. Eon Systems, embodied whole-brain emulation of a fruit fly, March 2026.
Contact
press@flylabs.dev. Logo and brain renders on request.