Fly Language Model
A little language model. 166,700 fly neurons.
Ask something curious and see what comes back.
Enter to send · Shift + Enter for a new lineA tiny model. Replies can be imperfect.
A little language.
A lot of wiring.
Language model + real fly wiring. An experiment, not a biological fly speaking.
Your words enter a small language model. The fly’s wiring shapes a signal that nudges what it says next.
How this little thing works
A frozen 1.2-billion-parameter language model (Liquid AI’s LFM2.5) supplies the language. A fixed mapping sends its token embeddings through the real MaleCNS v1.0 wiring diagram. A trained 278,528-parameter adapter reads the resulting activity and adjusts the next-token scores. That adjustment is capped to keep the readout’s influence modest.
The input mapping, rate equations and readout are engineered. The graph stays fixed. There are no biological language pathways, dopamine rewards or online memory updates here.
The graph affects the output. We compare it with the language model alone and a similarly sized adapter without fly wiring. Causal influence does not establish a biological advantage.
What happens to my chat?
Messages are processed on our server. We don’t save chat histories or train on your messages. The model has a short context, so older turns fall out as your conversation grows. Your conversation stays in this tab while you visit this page. Reloading the page starts a fresh chat.
Messages use shared server capacity. The waiting queue is limited, so you may be asked to try again soon when the fly is busy. Your message stays in the composer.