The Fly Language Model (FLM) has arrived, marketing itself as the world’s first chatbot powered by a complete fruit fly connectome. It couples the MaleCNS v1.0 graph—containing 166,700 nodes and over 25 million edges—with a frozen LiquidAI LFM2.5-1.2B backbone. The architecture, dubbed Generative Pre-trained Fly (GPF) by its creator, is designed to run locally on Python 3.12 without API keys. It sounds like a breakthrough in neuromorphic computing. The reality is far more mundane.
Only 278,528 parameters are actually trained in this setup, representing a tiny 0.0238% of the total model size. The system functions as a reservoir computer where the fly brain graph processes token embeddings before they hit the language model. On a test set of 32 SmolTalk dialogues, the FLM did improve perplexity from 3.98 to 3.90. That looks promising until you look at the controls. A simpler model that bypassed the fly graph entirely outperformed the FLM in every single test seed. The paired bootstrap interval confirms there is no statistically significant gain from using the fly topology.
The developer is transparent about these limitations, explicitly disclaiming that FLM is the first connectome language model. Previous work like ngxson/fly-hf used smaller subsets of the fly brain. The key distinction here is scale and the use of a frozen backbone to isolate the source of linguistic competence. Tests show the graph does participate; setting the connection weights to zero removes the effect entirely. However, relabeling the nodes breaks the performance, proving the readout depends on learned alignment rather than any magical property of fly wiring.
Crucially, the recurrent state forgets information rapidly, contracting differences by a factor of 0.6 per token. After just 20 tokens, the memory bound drops to near zero. This means the connectome adds no long-range memory capabilities; all context still comes from the standard transformer backbone. While the MIT-licensed code is deployable on macOS or Linux, the study artifacts remain private. Without independent reproducibility, this remains an interesting engineering footnote rather than a paradigm shift.
