A language model can appear to “use a calculator” while relying on software running outside the model. This experiment reverses that assumption: its arithmetic engine is compiled into added transformer layers, and the model’s own recurrent state stores the CPU’s working memory. The result explores whether exact computation can be part of inference itself, rather than an external tool call.
AI Model2026
Qwen3.5-4B-Qwenputer
Compiles a deterministic RISC-V CPU into a language model’s weights, enabling exact arithmetic and C or MicroPython execution without an external tool runtime. It uses Gated DeltaNet state as memory and remains experimental.
Introduction
Information
- Websitehuggingface.co
- AuthorsDivyansh Agrawal
- Published date2026/10/19
Categories
More Items
AI Model2026
Provides GGUF quantizations of DeepSeek-V4.1-Flash-1017 for local inference, with swappable Engram n-gram tables. Even the smallest quant requires about 128 GB of memory; SSD offloading can reduce the amount kept resident.
AI Model2026
Provides quantized Qwen-Image-2.1-Turbo weights for text-to-image generation across GGUF, Safetensors, and MLX formats. It retains the model’s 8-step Turbo workflow, with options for different hardware and runtimes.
AI Model2026
Reduces refusals by removing a refusal direction from selected dense residual writers in a quantized DeepSeek-V4.1-Flash-1017. It preserves the upstream GSQ-RCO routed experts and offers three low-bit size options.