Fly Language Model (FLM) Wires the Fruit Fly Connectome Into a Frozen 1.2B LLM — and Its Own Controls Show the

Overview

The Fly Language Model (FLM) is a publicly available chatbot that couples the complete retained MaleCNS v1.0 fruit fly connectome to a frozen LiquidAI LFM2.5-1.2B-Instruct backbone. Its developer describes it as the world's first Fly Language Model, built on an architecture dubbed GPF (Generative Pre-trained Fly). Notably, the project does not lean on the GPF label as a marketing claim, explicitly disclaims being the first connectome language model, and — most tellingly — reports that a parameter-matched control without the fly graph performs slightly better.

Deployable: Yes, locally. The nftechie/flm repository is MIT-licensed and runs on Python 3.12 (macOS or Linux; MPS, CUDA, or CPU) with no API key required.

What Was Actually Built

At its core, the system is a reservoir computer bolted onto a language model. All 166,700 retained nodes and 25,582,938 directed edges of the MaleCNS graph participate in the computation. The graph, the backbone, and the random input and output projections are all held fixed. Only a 278,528-parameter readout is trained — roughly 0.0238% of the 1,170,340,608 backbone parameters.

The mechanics work as follows:

• Input compression: At each token, a fixed Gaussian projection compresses the 2,048-dimensional token embedding down to 128 channels.
• Reservoir injection: Each reservoir node receives one channel with a random sign.
• Graph update: The entire graph updates according to x = tanh(W(0.6x + 0.4Bc)), where W holds incoming-normalized anatomical contact counts.
• State pooling: Reservoir states are pooled into 128 bins, passed through two trained bias-free matrices (U at 128×128, V at 2,048×128), and projected through the frozen vocabulary head.

via MarkTechPost