Models

Domain models for sensitive work

Open AI models built for one field each, like patents, security, law, money, health, and space math. Each comes with a benchmark, downloads free, and runs offline on hardware you control.

Flagship Model

A small model that answers from your own notes and files and names the exact source. When the answer is not there, it says so instead of making one up. It runs offline on hardware you control.

A movie-poster style picture. A dark noir records room of filing drawers, where a small gold AI computer shines a beam of light on one open drawer and the one page inside it, with a detective hat on the desk. The title reads The Source.

Advisory

Advisor 4B · GGUF

Answers from your own notes and files, with the exact source named. When the answer is not there, it says so instead of making one up. A small model you run on your own hardware.

Built on: nvidia/NVIDIA-Nemotron-3-Nano-4B
Format: GGUF
Best build: Q4_K_M
License: NVIDIA Nemotron Open Model License

Patent Strategist

This model reads and reasons about patents fully offline. Built with the NeMo toolkit, it comes in two formats: ready-to-run files, or a small add-on patch for builders.

A movie-poster style picture. A Victorian inventor workshop full of brass gears and blueprints, with a small gold AI computer on the desk. The title reads Prior Art.

Patent

Patent Strategist nemo · GGUF

Offline patent reasoning in ready-to-run files, built with the NeMo toolkit. Nothing leaves your desktop.

Built on: deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
Format: GGUF
Best build: Q5_K_M
License: Apache-2.0

Patent

Patent Strategist nemo · adapter

The NeMo-built patent reasoning as a small add-on patch for the base model. For your own custom builds.

Built on: deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
Format: LoRA adapter (BF16)
Best build: BF16
License: Apache-2.0

More domain models

Each one is tuned for a single field, so it knows the words and rules of that work.

A movie-poster style picture. A 1940s code room with old cipher machines, where a small gold AI computer turns secret code into a glowing shield. The title reads Zero Day.

Security

SecurityLLM

Tuned for cyber security questions, threat write-ups, and security know-how.

Built on: ZySec-AI/SecurityLLM
Format: GGUF
Best build: Q4_K_M
License: Apache-2.0

A movie-poster style picture. A grand courtroom with a gavel, where a small gold AI computer glows beside the case files. The title reads The Verdict.

Legal

Saul 7B Instruct

Tuned for legal text and built to follow instructions on legal tasks.

Built on: Equall/Saul-7B-Instruct-v1
Format: GGUF
License: MIT

A movie-poster style picture. A 1980s trading floor with a golden bull and rising green charts, with a small gold AI computer on the desk. The title reads The Bull Run.

Finance

Finance Chat

Tuned for finance and money questions in plain chat.

Built on: AdaptLLM/finance-chat
Format: GGUF
License: Open

Medical

II-Medical 8B

Tuned for medical questions and clinical text.

Built on: Intelligent-Internet/II-Medical-8B
Format: GGUF
Best build: Q5_K_M
License: Apache-2.0

A movie-poster style picture. An old observatory at night, with a brass telescope and golden orbit lines drawn across a starry sky, and a small gold AI computer on the desk. The title reads The Third Law.

Space

Kepler

Tuned for space math, like the paths satellites fly. It shows short work and ends in one number you can check.

Built on: Qwen/Qwen3-8B
Format: GGUF
Best build: Q8_0
License: Apache-2.0

Benchmark dataset

The test set we use to check how well a model reasons about patents.

A movie-poster style picture. An exam hall with a big score gauge and a chalkboard of grades, where a small gold AI computer marks the test papers. The title reads The Final Exam.

Dataset

Patent Strategist Bench

A 200-question test set that scores how well an AI reasons about patents, across seven kinds of task.

Format: Dataset · 200 questions
License: CC-BY-4.0

The technical details

All weights are open. Advisor is a 4B model built on NVIDIA Nemotron, tuned for grounded question-and-answer over a retrieval corpus with exact source-id citations and clean refusals on missing-source or private questions. Its promoted serving lane is the Q4_K_M GGUF, about 2.6 GB and roughly 70 tokens per second on Spark-class hardware, with bench behavior identical to Q8_0. The two Patent Strategist builds start from DeepSeek-R1-0528-Qwen3-8B, are made with the NeMo toolkit, and ship two ways: GGUF files (ready to run with llama.cpp) and LoRA adapters (small training patches in BF16). On Spark-class hardware (a small ~$3,000 AI desktop) the Q5_K_M GGUF runs around 32 to 35 tokens per second. SecurityLLM builds on ZySec-AI/SecurityLLM, Saul 7B on Equall/Saul-7B-Instruct-v1, Finance Chat on AdaptLLM/finance-chat, and II-Medical 8B on Intelligent-Internet/II-Medical-8B, each shipped as GGUF from Q4 through Q8 and F16, and scored on domain tests like CyberMetric, LegalBench, FinanceBench, and MedMCQA. The Patent Strategist Bench is 200 questions across seven shapes (claim drafting, prior-art ranking, MPEP-grounded reasoning, and more), released CC-BY-4.0.