Mini Sentinel µAI-32

A whole AI in 30 KB. A trainable neural classifier in plain C. The audited x86-64 Linux executable is 9,188 bytes and its MNIST model is 21,468 bytes: 30,656 bytes combined, under a 32,768-byte limit. That pair classifies 9,781 of the 10,000 standard MNIST test digits (97.81%) after 2×2 pooling to 14×14.

Try it: draw a digit in your browser.

For scale: our Sentinel S1-32B has 32 billion parameters. µAI-32's digit model has 9,946.

Files

file bytes what it is held-out result
dist/uai32 9,188 the program: train, test, predict (x86-64 Linux) —
dist/mnist14.model 21,468 digits, 196-48-10, 9,946 parameters 9,781/10,000 = 97.81%
dist/spirals.model 348 two spirals, 2-32-2 298/300 = 99.3%
dist/rings.model 222 three rings, 2-16-3 300/300 = 100%
dist/iris.model 174 iris flowers, 4-8-3 29/30 = 96.7% (one test row duplicates a training row)
dist/SHA256SUMS checksums of the five files above
uai32.c the full source (257 lines); build files and tests are on GitHub
uai32.js the browser demo's forward pass; same answers as the C program on every test set above
refcheck.py independent NumPy check of inference and one training step

Use

sha256sum -c dist/SHA256SUMS
cc -O2 uai32.c -o uai32 -lm             # or use dist/uai32 as is; the exact size-optimised build is in the GitHub Makefile
echo "0.5 -0.2" | ./dist/uai32 predict dist/spirals.model     # class, then each class's probability
./uai32 test mnist14_test.txt dist/mnist14.model             # MNIST as 14x14 text rows; see the GitHub repo's get_mnist.py

Model file format (little-endian): u16 magic 0xA132, u16 inputs, hidden, outputs; float32 mean and scale per input; then every weight as bfloat16, hidden layer then output layer (last column is the bias). One hidden ReLU layer, softmax output, trained by SGD.

Evidence and limits

The original focused audit (7 Oct 2026) reports PASS on 28/28 mandatory checks, including exact rebuilds, sanitizer runs and a scrambled-label control (47% on spirals, chance 50%). Full evidence, reproduction guide and the "Disprove us" challenge: GitHub.

This is a one-hidden-layer classifier, not a language model. The size figure counts the executable and one model file only. It excludes the host C library, operating system, RAM, datasets and tools. MNIST images are not redistributed here.

Credits

Project direction and release stewardship: Sentovara. Engineering contributors: OpenAI GPT-5.6 Sol, Anthropic Fable 5.1 and OpenAI Codex. Browser demo: Anthropic Claude. Code Apache-2.0; Iris data notice in THIRD_PARTY_NOTICES.md.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Space using Sentovara/Mini-Sentinel-uAI-32 1