qwenFS and qwenGPU: a filesystem and SIMT GPU compiled into Qwen3.5-2B's weights
CatAstro_Piyush · x · 2026-11-04
Developer TheVixhal presents two experiments turning Qwen3.5-2B into a self-contained "computer":
- qwenFS: adds 8 compiled layers on top of Qwen's 24, using attention to hold files and Gated DeltaNet state to hold the file tables. Writes store text in the copy layer's attention log keyed by file, version and offset; reads use the same keys so attention copies text back token by token. Everything runs inside the model's forward pass.
- qwenGPU: stacks 45 compiled layers implementing a SIMT GPU in the weights — 4-lane warps, a 4-stage pipeline, 15 registers per lane, 126 words of global memory, all held in DeltaNet state. The host compiles CUDA kernels to PTX, then the model generates each launch step (registers, branches, memory writes) token by token.
A proof-of-concept showing how attention and DeltaNet state can carry structured computation inside model weights.
More from Fun
- AI-and-fiction book gets its first negative review from a literary critic — _akpiper · 2026-11-04
- Great idea? It's already been built 117 times on Product Hunt in the last 10 seconds — Dan_Jeffries1 · 2026-11-04
- AI vs bureaucracy will be bloody: 30-minute work sitting in a 7-day queue — airkatakana · 2026-11-04
- NFS Most Wanted nostalgia: Blacklist #14 Taz resurfaces after all these years — CatAstro_Piyush · 2026-11-04
- Creator Turns a Band's Music Video Into 'Slopcore' With His Claude Plugin — AIandDesign · 2026-11-04
- Coding in the shower: dev mounts iPhone on bathroom wall to approve agent tasks — TheMoonMidas · 2026-11-04