Numbat: A From-Scratch AI Engine That Drops Hundreds of Dependencies
Could mean AI that's cheaper to run, easier to ship, and safer in hospitals.
Almost all modern AI is built on Python plus hundreds of add-on packages that must all match versions. It works, but it's costly: installations break, you need separate tools to ship the finished model, and the code researchers write isn't the code products actually run. Numbat is a ground-up rebuild that avoids all of that. Everything is written in one general-purpose language, Zig, with no third-party runtime dependencies — nothing to install, nothing to version-match.
Despite being self-contained, it covers the full pipeline: tensor computation (the grids of numbers AI calculates with), automatic differentiation (how a model learns from its mistakes), neural-network building blocks, mixed precision, multi-GPU training, data loading and monitoring. An interface exposes more than 1,400 functions that other programs can call, with bindings for six programming languages. A clinical version even turns hospital regulatory rules into automatic pass/fail checks instead of written documentation.
Verifying the system was the harder half. A broken AI training run rarely crashes — it just quietly produces a slightly worse model, and nobody notices. So the team treated a widely used reference implementation as the benchmark and checked against it at five levels, from basic math up to full training runs. That caught ten 'silent divergences': places where the recipe secretly differed in ways that would have degraded results without any error message.
As the final test, they trained an image detector of the popular YOLO family (25.9 million settings) from scratch on 5,500 hours of labeled photos, for the full 500-round schedule. Its accuracy scored 0.4956 versus the reference's published 0.502 — effectively a tie — at identical speed on the same hardware. The weights, per-round metrics and full run record were released publicly for anyone to check.
- Numbat rebuilds the entire AI toolkit in one language with zero outside packages, so there's no fragile chain of installs to break.
- It matched a standard reference model's accuracy (0.4956 vs 0.502) while running at the same speed on identical hardware.
- The team uncovered 10 hidden differences from the standard version — exactly the kind of quiet mismatch that silently makes AI worse without any error.
Why It Matters
Fewer broken AI setups, easier shipping to phones and hospital gear, and a way to prove AI actually works.