Qwenputer

Qwenputer is an experimental model by developer Divyansh Agrawal that compiles a 32-bit RISC-V CPU directly into the weights of Qwen3.5-4B. It is not a plugin, tool call or sandbox: the CPU is 11 transformer layers placed in front of Qwen's own 32, and none of their weights are learned; a compiler computes every one. On 2026-10-28 the project became an X trending news story.

Model: huggingface.co/divagr1925/Qwen3.5-4B-Qwenputer · Code: github.com/divagr18/qwenputer · There is also a Qwenputer-Filesystem variant with a persistent in-model filesystem · Model license Apache-2.0.

Key points

1. Gated DeltaNet as RAM: Qwen3.5's DeltaNet layers carry a running state that the delta rule can overwrite one row at a time, exactly, so each row works as a memory cell. The CPU keeps its code, data, registers and program counter there.
2. Capabilities: runs C and MicroPython programs inside the model's forward passes and passes the official riscv-tests suites; the ISA is RV32IM + Zba plus a bignum coprocessor.
3. How it's used in chat: it piggybacks on Qwen3.5's tool-call format. Qwen writes a <tool_call> for calc, but nothing outside the model runs it; the computer layers read the call, execute it on the internal CPU, and the model itself writes the <tool_response> digit by digit.
4. Benchmarks (greedy, 40 problems per set): synthetic integer arithmetic 75.0 → 97.5; DeepMind mul_big 55.0 → 97.5; div_big 72.5 → 97.5; mul_div_multiple_longer 90.0 → 100, on par with base Qwen calling a real Python tool. About 230,000 replayed output positions showed zero mismatches against a reference machine, and text behaviour matches the base model (KL 6e-4).
5. Costs: 18.8 GB in bf16 (mostly zero padding; the computer has about 1M nonzero weights); the Triton executor runs about 3,200 trace tokens per second on an L4, slower than native hardware.

Why it's tracked

A distinctive name and a novel idea ("a computer compiled into LLM weights") that became an X trending story in this scan window. The weights were created on Oct 19, but the radar had not tracked it before.

Quick start

git clone https://github.com/divagr18/qwenputer && cd qwenputer/experimental/riscv && uv sync --extra hf
hf download divagr1925/Qwen3.5-4B-Qwenputer --local-dir qwenputer-rv
uv run python -m chipformer.rv.chat qwenputer-rv --firmware tests/data/rv/calc-bridge.elf \
    --ask "Use calc: what is 123456789 * 987654321?"

Tips

• The model loads as a stock Qwen3_5ForCausalLM; the RvChat driver loads firmware into the DeltaNet state and manages trace tokens.
• It is a research experiment for exploring exact computation inside a model, not a production calculator.
• Community FP8 quantizations already exist (e.g. liodon-ai/Qwen3.5-4B-Qwenputer-FP8).