Oriel — AI & technology worth noticing, every day.

divagr1925/Qwen3.5-4B-Qwenputer

AI model · https://huggingface.co/divagr1925/Qwen3.5-4B-Qwenputer

First seen by Oriel on .

AI summary

Qwenputer adds a compiled 32-bit RISC-V CPU to Qwen3.5-4B’s weights, enabling exact internal code execution while largely preserving the base model’s text behavior.

  • The CPU comprises 11 non-learned transformer layers placed before Qwen’s 32 layers and uses Gated DeltaNet state as RAM for code, data, registers, and the program counter.
  • It runs C and MicroPython inside model forward passes, passes the official riscv-tests suites, and executes tool-call-formatted requests without an external runtime or tool executor.
  • Across five arithmetic benchmarks with 40 problems per set, Qwenputer scored 97.5–100%, compared with 55–90% for base Qwen3.5-4B; host Python scored 95–100%.
  • A replay of about 230,000 computer output positions found zero mismatches, while text-behavior KL was 6e-4 against a 5e-4 noise floor.
  • The model uses RV32IM with Zba and a bignum coprocessor, has 230,324 added trace tokens, and is 18.8 GB in bf16, mostly zero padding.

#qwen3.5 #risc-v #transformer models #deltanet

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 171, latest observation ; +69 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 1,603, latest observation ; +1,031 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Modelsdivagr1925/Qwen3.5-4B-Qwenputerhttps://huggingface.co/divagr1925/Qwen3.5-4B-Qwenputer Read discussion