~80KB
Unit tests for AI models. 80KB. Zero installation.
CI/CD model validator that runs real inference without Python or PyTorch.
LinuxThe problem
You push a new model — a fine-tune, a quantization, a merge. How do you know it is not corrupted? That the quantization did not destroy quality? That the fine-tune did not introduce regressions? Today: you cannot validate models in CI. Installing PyTorch in GitHub Actions takes minutes and 4GB. There is no model equivalent of unit tests.
The solution
OmniValidate is an 80KB binary that reads a GGUF model, scans for NaN/Inf in every tensor, runs actual inference with test prompts, and compares outputs against a golden baseline. Drop it into any CI pipeline. Zero installation. Zero dependencies.
Why Bare-Metal Matters
Model validation requires running inference. Running inference normally requires PyTorch (4GB+) or llama.cpp (compilation + dependencies). OmniValidate contains a full transformer inference engine in 80KB because OmniOS compiles directly to syscalls. This makes model testing as fast and simple as running a linter.
Technical Specifications
| Feature | Value |
|---|---|
| Binary Size | ~80KB |
| Function | CI/CD model validator with real inference |
| Formats | GGUF (Q4_K, Q6_K, Q8_0, Q5_0, F16, F32) |
| Dependencies | None — no Python, no PyTorch, no CUDA |
| Checks | Integrity, NaN/Inf, inference regression test |
| CI Ready | wget + chmod + run — zero installation |
Comparison
| OmniValidate | Python + PyTorch | llama.cpp | |
|---|---|---|---|
| Size | ~80KB | 4GB+ (torch + transformers) | ~2MB (compiled) |
| CI installation | wget (80KB) | pip install (minutes) | cmake + make (minutes) |
| Runs inference | Yes (full transformer) | Yes | Yes |
| NaN/Inf scan | Built-in | Custom script needed | No |
| Dependencies | None | Python, CUDA, numpy | libc, libstdc++ |
| Works in FROM scratch | Yes | No | No |
Use Cases
CI Model Gate
Add to your GitHub Actions or GitLab CI. Every model push is validated with real inference before merge. Catch corrupted quantizations, NaN explosions, and quality regressions automatically.
Quantization QA
After quantizing a model to Q4_K, run OmniValidate to verify the output has not degraded beyond acceptable thresholds.
Fine-tune Regression
Compare your fine-tuned model against the base model on a fixed prompt set. Detect if the fine-tune improved the target domain without breaking general capability.
Try Now — Free
Coming Soon
This product is under active development. Contact us for early access or to be notified when binaries are available.
Talk to the Team