The Fly Language Model (FLM) is a public chatbot that connects the full MaleCNS v1.0 fruit fly connectome to a frozen LiquidAI LFM2.5-1.2B-Instruct backbone. The developer, who calls it the world’s first Fly Language Model using the GPF architecture, explicitly disclaims being the first connectome language model and reports that a parameter-matched control without the fly graph performs slightly better.
Deployment details
The nftechie/flm repository is MIT-licensed and runs on Python 3.12 for macOS or Linux, supporting MPS, CUDA, or CPU with no API key required.
System architecture
The system functions as a reservoir computer attached to a language model. All 166,700 retained nodes and 25,582,938 directed edges of the MaleCNS graph participate. The graph, the backbone, and the random input and output projections remain fixed. Only a 278,528-parameter readout is trained, which accounts for roughly 0.0238% of the 1,170,340,608 backbone parameters.
At each token, a fixed Gaussian projection compresses the 2,048-dimensional token embedding to 128 channels. Each reservoir node receives one channel with a random sign. The whole graph then updates with x = tanh(W(0.6x + 0.4Bc)), where W holds incoming-normalized anatomical contact counts. States are pooled into 128 bins, passed through two trained bias-free matrices (U at 128 by 128, V at 2,048 by 128), and projected through the frozen vocabulary head as a bounded residual added to the backbone logits. The residual is capped at an RMS of 0.25 across vocabulary coordinates.
Performance results
On a freshly frozen set of 32 SmolTalk everyday-conversation dialogues containing 1,236 target tokens, three fit seeds produced the following results:
| Condition | NLL (nats/token) |
|---|---|
| Frozen backbone | 1.381995 |
| Fly readout | 1.359816 ± 0.000110 |
| Direct-input readout | 1.359328 ± 0.000108 |
| Relabeled, no refit | 1.381265 ± 0.000802 |
| No edges | 1.381995 |
The fly readout improved on the backbone by 0.0222 nats per token, lowering perplexity from 3.98 to 3.90. However, a direct-input control, which feeds the same 128-channel token projection straight into an identical readout with no graph, performed better in all 3 seeds by 0.000488 nats per token. The paired bootstrap interval (+0.00000502 to +0.00104) does not support a fly-specific gain.
Two other controls matter. Setting W to zero removes the residual exactly, reproducing the backbone’s per-token losses, so the graph verifiably participates. Relabeling node identities without retraining returns NLL near baseline, which shows the readout depends on its learned interface alignment, not that fly topology beats random wiring.
The research report also proves the recurrence contracts initial-state differences by at most 0.6 per token. After 10 tokens that bound is 0.00605; after 20 it is 0.0000366. Piling in 166,700 cells does not buy long memory. Context still comes from the backbone.
Prior work and the first claim
The research report cites ngxson/fly-hf, an earlier prototype that used a 49,393-cell central-brain subset of MaleCNS as a reservoir trained on TinyStories without a pretrained backbone, and states plainly that it makes no claim to be the first connectome-based language model. FLM’s distinction is scale, using the full retained graph, and the frozen-backbone design that keeps the source of language competence identifiable.
What it means
For people building tools, the experiment adds a specific data set to the mix but offers no immediate performance boost. The neural wiring of the fruit fly does not improve text generation over a standard random projection when attached to a frozen large language model. The connectome participates in the calculation but does not provide a functional advantage for the output quality.
