library

Noir

A lightweight Transformer framework built entirely in C. Zero dependencies, transparent ops, no abstractions.

2025-12-30ยทseedling#projects#c#transformers#systems

Noir is a lightweight Transformer framework built entirely in C. Think Darknet, but designed specifically for the Transformer architecture.

if you don't understand what it's doing at the byte level, you don't understand it.

Why Noir?

Most modern transformer implementations are bloated with abstractions. Noir aims to be:

  • Zero-Dependency: No heavy libraries, just pure C and standard linear algebra.
  • Transparent: Every tensor operation, every attention head, and every residual connection is clearly visible and tuneable.
  • Fast: Specialized kernels for MatMul and Softmax, optimized for CPU and eventually GPU.

Objectives for 2025

  1. Implementation of foundational Multi-Head Attention blocks.
  2. Custom C-based tensor serialization format.
  3. Training a small-scale model (under 100M parameters) from scratch.

Timeline

This is a live feed of progress on Noir.

2025-12-30: Project Inception

  • Defined core architecture principles.
  • Initializing the repository.
  • Designing the Tensor struct.

"To understand the transformer, one must build it from the byte up."

2025-12-25: Early Research

  • Skimmed through the "Attention Is All You Need" paper for the 100th time.
  • Looking into C-native BLAS libraries vs writing my own kernels.