Reaction
qwenFS is presented as a filesystem built into Qwen3.5-2B's weights
qwenFS's introduction says its eight extra layers use attention to hold files and Gated DeltaNet state to hold tables.
TLDR
qwenFS is pitched as performing saves and reads inside the model's forward pass. Writes store text in an attention log keyed by file, version and offset; reads use those keys to copy it back token by token. One X user called the idea neat but questioned whether putting a whole filesystem inside an AI model makes sense.
Combined views
—
3 Sources, first seen ago
