We train, release, and run specialised AI models you own.

naderu-geek-py-0.5b

Our first public portfolio model: a compact LoRA fine-tune of Qwen2.5-Coder-0.5B-Instruct (Apache-2.0) that imprints a clean, commented-Python style and a recognisable naderu-geek identity. It demonstrates our pipeline — data → training → evaluation → release — end to end. It is a demonstration piece, not a capability claim: coding ability tracks the base. It passed a qualitative smoke eval (identity recognised · coherent Python), with the card and provenance stated plainly.

naderu-loom-7b

An offline interactive-fiction narrator: given a game state and the player's action, it returns a single JSON turn (scene + 2–4 choices + state changes) an app can render and apply — a game master that runs on-device. A QLoRA fine-tune of Mistral-7B-Instruct-v0.3 (Apache-2.0), trained on Apple Silicon. Unlike a smoke test, it ships a quantitative eval gate — and passes it: 100% valid JSON and 0 world-state violations across 34 scripted turns.

hunter-crypto-7b

The first model of the Hunter family — a local cryptography-attack specialist for authorized security and CTF work. Given a weak or misconfigured construction, it names the flaw and emits a runnable attack script that recovers the secret, fully offline on Apple Silicon. A QLoRA fine-tune of Qwen2.5-Coder-7B-Instruct (Apache-2.0). Its gate is execution-graded — the attack recovers the secret or it does not — and it passes: 100% solve rate on the in-distribution held-out set and 97% on an independent-method set. A narrow specialist, authorized use only.

naderu-mini-4b

Live in production inside AiToolK.it — the offline, on-device fallback the app switches to with no connection: rewrite, summarize, prompt-smithing, everyday Q&A, all generated locally, no cloud call, no data leaving the phone. An identity/behaviour fine-tune of Qwen3-4B (Apache-2.0), trained on Apple Silicon. It ships a quantitative behaviour gate (8 metrics, held-out probes) — and passes all eight: it stays in character, stays honest about being offline (declines live-data tasks, never fakes app actions), and names its Qwen3 base plainly even under adversarial probing — a rebrand that never hides provenance. The gate is run on the shipped GGUF itself.

naderu-laya-150m

A small, fast decision router: it sorts an incoming message into billing, technical support, sales, cancellation, or a question about the model itself — in about 11 ms on Apple Silicon and 10 ms on CPU with ONNX Runtime. A LoRA fine-tune of ModernBERT-base (Apache-2.0) with calibrated confidence. Its gate tests phrasings it never saw in training: 93.6% top-1 with calibration error 0.036. Our first cut scored 100% — until we stopped letting test sentences share templates with training, so the number we publish is the honest one.

Restrictions trained into the weights.

Hand us a policy spec — blocked domains, forbidden ports, commands your agent must never run — and we deliver a model tuned to observe it, plus the adversarial eval that measures adherence: refusal rate on forbidden actions across injection, roleplay, and encoding attacks — and over-refusal rate on allowed ones, held to the same bar. Built for regulated, on-prem, and air-gapped agents, where prompted guardrails don't travel with the model.

Honest scope — fine-tuning yields a behavioral tendency, not a guarantee. This is defense-in-depth: every delivery pairs the tuned model with a stated harness-level enforcement recommendation. We don't promise "no matter what" — we promise a measured number.