Qwenputer front panel

Qwen3.5-4B with a 32-bit RISC-V CPU compiled into its weights. When Qwen calls a tool, that CPU runs it inside the model and writes the answer.

CPU speed
0.0 s
Qwen writes
CPU writes
0x00000000PC
0instructions
0CPU tokens
–real clock

Layers a CPU token runs layers 0–10 only; a text token runs all 43

0–10 compiled CPU11–42 Qwen3.5-4B, unchanged

Registers x0–x31, one light per bit

Code

RAM 32 KB, one pixel per word

0x0000 datastack 0x7fff

Digit file the bignum unit: 8 registers of decimal digits, most significant on the left

Tool-call buffer

CPU tokens

What you are watching

Qwenputer puts 11 compiled layers in front of Qwen3.5-4B's 32. A compiler set every weight in them, none were trained, and Qwen's own weights are unchanged. Together they are an RV32IM CPU with a bignum unit. Its program, registers and 32 KB of RAM are held in the layers' Gated DeltaNet state.

How a tool call runs

Qwen's tool-call tokens land in a buffer the CPU reads at address 0x8000. At </tool_call> the CPU takes over and runs toolbox.c, which picks the tool by name. It writes one to a few tokens per instruction and prints the answer as ordinary text. Its tokens skip Qwen's layers, so Qwen only sees the result. The CPU reads the bytes of Qwen's tokens, which is how it can count letters Qwen cannot see.

About the recordings

Each run was recorded on an NVIDIA A10G with greedy decoding. A reference RISC-V machine ran in lockstep and had to produce every CPU token the model wrote, and the panel shows that machine's registers and memory. The CPU runs one to two thousand instructions a second, so short runs play the CPU at 1/10 speed.