Gadeon

Gadeon runs Qwen3.5 and QwenPaw hybrid language models entirely on your device's Apple Neural Engine, using Core ML as its only framework dependency. There's no C, no Python, and no cloud calls: every token is generated locally, keeping your conversations private. Under the hood, these models mix Gated DeltaNet recurrent layers with a smaller number of attention layers. Gadeon compiles that hybrid architecture directly onto the Neural Engine's dataflow accelerator, so both prompt processing (prefill) and token generation (decode) run on-chip instead of falling back to the CPU or GPU. On an Apple Silicon Mac, prefill measures roughly 2 to 2.4 times faster and about 4.4 times more energy-efficient per token than a Metal GPU baseline, while leaving the GPU free for other work. Conversations in Gadeon are append-only: each turn extends the model's fixed-size recurrent state instead of re-processing the entire chat history, so response latency doesn't grow as a conversation gets longer. The app can also roll back to an earlier point in a conversation, dropping reasoning traces or replaying tool calls in a clean form without reprocessing everything from scratch. For ternary-quantized models, Gadeon switches to a pure-Swift Metal GPU backend behind the same session and tokenizer interface, so both compute paths feel identical from the chat UI. Gadeon is a proof of concept for bringing Gated DeltaNet hybrid models to the Neural Engine through Core ML - built to explore what's possible on Apple's own silicon.