divagr1925/QwenPU

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

QwenPU is Qwen3.5-4B with a 32-bit RISC-V CPU and a vector unit compiled into its weights. The vector unit runs a whole neural network inside the model's forward passes: Andrej Karpathy's stories260K, the 260K-parameter Llama from llama2.c, runs as a program on it and writes its stories.

It's part of Qwenputer: the same CPU, compiled into transformer layers in front of Qwen's own, every weight set by a compiler, with Qwen3.5's Gated DeltaNet state as its memory. Qwenputer-npu gave that CPU dot-product instructions. QwenPU gives it a vector unit of its own: separate memories in DeltaNet heads, 64-lane matrix rows a token, and whole vector results streamed through the token stream.

Work in progress. QwenPU is exact, but not yet small: the checkpoint is 26.7 GB, so it needs a 40 GB or larger GPU to load whole. The story demo and the checks run the computer's layers on their own and fit a 24 GB GPU. Trimming the hidden size, and with it the checkpoint, is next.

A language model inside the language model

stories260K is a C program here (guests/stories260k/stories_vx.c). Its int4 weights sit in the vector unit's weight memory, loaded from the program, and every matrix row, residual add, softmax and attention read is a vector op. None of it is compiled into the checkpoint, so any model that fits runs the same way.

Every build writes the same story, token for token:

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

The vector unit

Its memories are DeltaNet heads: 16 vector registers of 64 words, 384 int8 activations, 4,152 row-chunks of int4 weights, and K / V caches for 32 positions. An op's record starts with config tokens carrying its operands, then streams its results, and each token is one exact step of the compiled layers. SiLU and exp are fixed-point function units, generic like the CPU's adders.

Checks

• Every vector op runs random test programs on this checkpoint's own chip layers, teacher-forced against a reference model of the unit, along with the first 30,000 steps of the story program: zero mismatches.
• The circuit's executor writes the whole story free-running on a GPU, checked against the reference at every one of its 3,184,920 steps: exact. It runs at about 3,300 trace tokens a second on an L4, so a story token takes about 16 seconds, against 41 for Qwenputer-npu.

Quickstart

git clone https://github.com/divagr18/QwenPU && cd QwenPU && uv sync --extra hf
uv run python guests/stories260k/storyrun.py guests/stories260k/stories_vx.elf --vx
hf download divagr1925/QwenPU --local-dir qwenpu
uv run python -m chipformer.rv.vxcheck qwenpu guests/stories260k/stories_vx.elf

storyrun writes the story on the computer's executor; vxcheck runs the vector unit on this checkpoint's layers (--chip-only loads only those on the GPU, about 13 GB).

Model details

Citation

Author: Divyansh Agrawal (ORCID: 0009-0003-0739-2473).

Cite this Hugging Face model artifact using its DOI:

@misc{agrawal2026qwenpu_model,
  author    = {Agrawal, Divyansh},
  title     = {{QwenPU}},
  year      = {2026},
  publisher = {Hugging Face},
  version   = {0ea39d7},
  doi       = {10.57967/hf/11460},
  url       = {https://huggingface.co/divagr1925/QwenPU},
  license   = {Apache-2.0},
  note      = {Model artifact. Author ORCID: https://orcid.org/0009-0003-0739-2473}
}

This DOI identifies the model artifact, not a GitHub source-code release. For the MIT-licensed source code, use the separate software citation in the QwenPU repository.

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 and llama2.c are by Andrej Karpathy (MIT).