QwenBend: a 53M-weight CPU that type-checks and runs Bend code inside a model's forward pass
Affectionate_Hat_585 · reddit · 2026-11-02
A developer released QwenBend, an experimental project where Qwen writes a Bend program, then the same model type-checks it with Bend2 rules and executes it inside its own forward pass — no interpreter, compiler, or tool calls. The entire CPU is 53M hand-set non-zero weights inside the model.
How it works
- After Qwen emits the Bend program, at <|imend|> the engine takes over and every next token is one executed instruction
- Hardware mapping: a 16-register CPU with 8K×32-bit RAM built from Gated DeltaNet; logic implemented with SwiGLU units acting as AND gates
- Pipeline: tokens → lexer builds bend2 lexemes → compile pass on the bendlite subset checking types, affinity, and termination → if no violations, evaluation runs on a return stack + operand stack
The author calls it the next step after Qwenputer (running an LLM inside an LLM), done "for lols" — but the finding that Gated DeltaNet + SwiGLU can constitute a whole computer inside a model is an interesting hook for formal verification work.
More from Fun
- Not true, I'm still at Ricursive: iScienceLuvr debunks departure rumor with a meme — iScienceLuvr · 2026-11-02
- Slop Eisley 0.6.80 patch fixes keyboard input bleeding into flight controls — jasonkneen · 2026-11-02
- Grok Confirms Beth Barnes' Deep Effective Altruism Ties, Prompting Debate Over Who Shapes Claude's Guardrails — adamamcbride · 2026-11-02
- AI psychosis: fully shaping your information environment around your own mental models — generativist · 2026-11-02
- fofrAI: vibe coding is just coding now — fofrAI · 2026-11-02
- Spotted: a door with a triangle logo that looks exactly like Anthropic's — yenkel · 2026-11-02