Projects neural systems

Loom

Analytical neural computer architecture for executing compiled programs inside looped transformers.

Role
Creator and maintainer
Paper
arXiv preprint, 2026
Testing
140+ regression tests
A bubble sort in C compiled to 49 Loom instructions; the real 155 by 1024 state tensor with the program counter and the current instruction highlighted; the eight fixed-weight layers that execute one instruction per forward pass; the instruction executed at each of the 486 forward passes; and the three model sizes
A real run of the bubble sort demo on the 155 × 1024 model. The C program compiles to 49 instructions from the 21-opcode ISA, held in the state tensor, shown here after 16 forward passes with the program counter at SUB 37, 36. Each pass through the eight fixed-weight layers executes one instruction, and the sort finishes after 486 passes.

Loom explores a training-free neural computer built as a looped transformer with analytically derived weights that can execute compiled programs inside a fixed architecture.

The public release connects systems research to usable artifacts. It includes a C-to-ISA compilation path, ONNX exports, Hugging Face model assets, WebGPU browser demos, FPGA verification, and a large regression test suite for validating programs under the same fixed-weight computational model.

Project details

Artifacts
GitHub implementation, browser demos, Hugging Face ONNX models, FPGA verification artifacts, and 140+ tests
Keywords
  • Analytical weights
  • Looped transformers
  • Neural computers
  • C compilation
  • ONNX Runtime WebGPU
  • FPGA verification
  • Client-side demos

References

2026

  1. Loom: A Scalable Analytical Neural Computer Architecture
    Loom: A Scalable Analytical Neural Computer Architecture
    Mehmet Kerem Turkcan
    Apr 2026
    arXivDOIHTMLCodeModel