# QwenBend : Qwen writes a Bend program, then the same model type checks it with Bend2 rules and runs it inside its own forward pass — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/a62bbdb1b6b9a3ab

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-01 20:58 ingested 2026-11-02 02:41

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xle8ta/qwenbend_qwen_writes_a_bend_program_then_the_same/)

## retained text



```
This is next step to Qwenputer perhaps the step that shouldn't have been taken but here it goes

Qwen writes a [Bend](https://bend2.dev/notes/what-is-bend2/) program, then the same model type checks it with Bend2 rules and runs it inside its own forward pass
There is no interpreter, compiler or tool call just like qwenputer. The whole CPU is 53M non zero hand-set weights inside the model

To see how it works: https://github.com/voidash/qwenbend/blob/main/docs/HOW_IT_WORKS.md

TLDR;

Qwen writes the bend program and at <|im_end|>, the engine takes over and every next token is one executed instruction

16 register CPU with 8k *32 bit
RAM = Gated DeltaNet 
Logic = SwiGLU unit as AND gates

1. You get your tokens first 
2. lexer walks and builds bend2 lexemes 
3. compile pass on subset of bend2 (bendlite) rules which checks types, affine and termination
4. if violations nothing runs else we run evaluations which is built using a return stack and operand stack


Why? 
For Lols... There is absolutely no reason. However Gated Deltanet with swiGLU allows for whole computer to exist inside a model. the original guy has built Qwenputer, run LLM inside LLM and i thought formal verification  


```



document fetched 2026-11-02 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1xle8ta/qwenbend_qwen_writes_a_bend_program_then_the_same
