Raven-1

Frontier-grade

The RavenSec family of first-party security models — purpose-built to find vulnerabilities, write fixes and tests, and reason about threats with frontier-grade quality. Every output is validated(it must compile, pass its tests, and re-scan clean) and every human decision makes the next model better.

Models ready · pipeline functional

The corpus, training (QLoRA + DPO), the correctness-gated evaluation harness, and optimized GPU serving are built and working — every fix is validated before it ships. The models are ready to train on enterprise GPUs, awaiting GPU provisioning.

Training signals↑ 23%

552

human-verified labels

Approved fixes

412

positive examples (SFT)

Corrections

140

rejected / edited

Preference pairs

140

for DPO training

Models

ModelSpecialisationStatus

raven-1-72bflagship · generalistready to train

raven-1-code-32bcode specialistready to train

raven-1-sec-72bsecurity specialistready to train

The flywheel

Agent proposes fix→Human approves / rejects→Labelled into corpus→SFT + DPO training (QLoRA, GPU cluster)→Eval beats baseline→Promote → optimized GPU serving→Serves the next fix

Pipeline built and functional today — corpus, QLoRA/DPO training, the evaluation gate, and optimized GPU serving. Ready to train on enterprise GPUs(FP8 mixed precision) the moment GPUs are provisioned.