divagr1925/Qwenputer-npu

Code on GitHub · Qwenputer · Base model: Qwen/Qwen3.5-4B · Apache-2.0

Qwenputer-npu is Qwen3.5-4B with a 32-bit RISC-V CPU compiled into its weights, and that CPU now has a neural-network instruction set. It is fast enough to run a second language model: Karpathy's stories260K runs as a C program on the CPU inside Qwen and writes its stories.

Like Qwenputer, the CPU is 11 transformer layers in front of Qwen's own, every weight set by a compiler, with Qwen3.5's Gated DeltaNet state as its RAM. This version adds two instructions and a weight memory:

• VDOT rd, rs1, rs2: the dot product of 8 int8 activations (two registers) with the next 8 int4 weights of the weight memory, added to rd. WB rs1 sets where the weights start.
• SMAQA rd, rs1, rs2: four signed byte products added to rd, as in the RISC-V P extension.
• A weight memory of 32,640 words (261,120 int4 weights) in DeltaNet heads, loaded from the program's .wrom section.

Each is one instruction for the CPU and one trace token, computed exactly by the compiled layers.

A language model inside the language model

stories260K is a 260K-parameter Llama trained on TinyStories. Here it is a C program (guests/stories260k/): its weights are data in the CPU's memory, and every multiply is an instruction the CPU executes. None of it is compiled into the checkpoint, so any other model that fits the memory runs the same way.

The int4 weights come from distilling the float model on its own samples. The CPU writes the same story as the float model for all 60 tokens checked:

Once upon a time, there was a little girl named Lily. She loved to play outside in the park. One day, she saw a big, red ball. She wanted to play with it, but it was too high

A new fused executor runs the CPU at about 5,500 trace tokens a second on an L4, so a story token takes about 41 seconds.

Checks

• The checkpoint's own chip layers run VDOT and SMAQA test programs and the first 30,000 steps of the story program, teacher-forced against a reference model of the CPU: zero mismatches.
• The executor runs the story free, checked against the reference: exact.
• An IR replay of 39 of the mul_big conversations (114,736 positions, 93,381 of them the computer's) against the reference: zero mismatches.
• calc still works the same way: every answer the CPU gave in these runs matched the reference model. The published sets, 40 problems each, batched decoding:

The four arithmetic misses are Qwen repeating calc calls until it runs out of budget, with the CPU answering each one correctly; the published Qwenputer scores 97.5 on that set. Qwen's text behaviour matches base: KL 5.8e-4, against a noise floor of 5.1e-4.

Quickstart

git clone -b npu https://github.com/divagr18/qwenputer && cd qwenputer/experimental/riscv && uv sync --extra hf
hf download divagr1925/Qwenputer-npu --local-dir qwenputer-npu
uv run python -m chipformer.rv.chat qwenputer-npu --firmware tests/data/rv/calc-bridge.elf \
    --ask "Use calc: what is 123456789 * 987654321?"
uv run python -m chipformer.rv.npucheck qwenputer-npu guests/stories260k/stories260k.elf
uv run python guests/stories260k/storyrun.py guests/stories260k/stories260k.elf

npucheck runs the new instructions on this checkpoint's layers; storyrun writes the story on the CPU's executor.

Gated DeltaNet as RAM

Model details

Citation

@software{agrawal2026qwenputer,
  author  = {Agrawal, Divyansh},
  title   = {{Qwenputer}: a computer compiled into the weights of a language model},
  year    = {2026},
  version = {0.3.0},
  url     = {https://github.com/divagr18/qwenputer},
  license = {MIT}
}

License

Apache-2.0. This is a modified version of Qwen/Qwen3.5-4B: the config is widened and computer layers are added, but Qwen's own weights are unchanged. stories260K is by Andrej Karpathy (MIT).