AI research, run on one desk.
The NVIDIA DGX Spark is Orionfold's reference machine. The book, open models, and tools below were all proven on it first.
Arena Field Edition for DGX Spark
Get an AI team without hiring one, in the Spark you own.
One installer turns your NVIDIA DGX Spark into a private AI lab with an AI research teammate, and proves itself on first boot. The proof of what it unlocks: one builder, solo on this exact stack, shipped 14 tools, 6 open models, 3 books, and 2 production sites. The software stays free and open. What you buy is the assembled, proven team, delivered turnkey.
Program membership
Orionfold is a member of the NVIDIA Inception program. The membership supports the same practical goal as this page: make private AI work easier to build, test, and run on customer-controlled hardware.
Offer fit on the NVIDIA stack
Orionfold Advisor
ShippingA governed local AI advisor over an enterprise corpus for small teams adopting local AI. The first Advisor shipped in June 2026: a fine-tuned NVIDIA-native 4B model lane with exact source-id citations, adversarially tested refusals, recall-gated retrieval, and deterministic frontier routing under a spend cap.
NVIDIA used: NVIDIA DGX Spark reference hardware, CUDA-capable local inference stacks
Private Agent Starter Kit
Alpha/BetaA fixed-scope starter kit that gives a small team one private agent workflow with trigger, runner, evaluation, approval loop, artifact output, and audit trail. It turns AI experimentation into an owned operating capability.
NVIDIA used: NVIDIA DGX Spark reference hardware, CUDA-capable inference runtimes
Orionfold Domain Experts
ShippingOffline model packs for sensitive domain work. Each pack combines a domain model, benchmark or scorecard, local run path, and use-case playbook. Patent Strategist is the live flagship; security, legal, finance, and medical packs follow the same pattern.
NVIDIA used: NVIDIA DGX Spark, NVIDIA NeMo, CUDA-capable inference
Orionfold Local AI Cockpit
ShippingA local cockpit for comparing, remembering, evaluating, and shipping private AI work. Arena provides the cockpit, Cortex adds local memory and provenance, and fieldkit carries reusable patterns for model runs, evals, RAG, training, and publishing.
NVIDIA used: NVIDIA DGX Spark, NVIDIA NIM, CUDA-capable inference, NVIDIA NeMo patterns
Most powerful AI lives far away, in a rented cloud. Orionfold builds the other way. Our reference machine is the NVIDIA DGX Spark, a small AI supercomputer that sits on one desk.
It is tiny but mighty. Inside is NVIDIA's GB10 Grace Blackwell chip and 128 GB of shared memory. That is enough to fine-tune (retrain) models up to about 70 billion parameters, and to run models up to about 200 billion, all on your own desk with no cloud and no meter running.
Everything below was proven on this machine first. The book is the field log of the work. The models are tuned and tested on it, with the flagship built using NVIDIA's NeMo toolkit (NVIDIA's open kit for training models). The fieldkit toolbox is the set of patterns that held up on it.
The Spark is our reference, not our only home. The same models and tools also run on Apple Silicon and other small devices, a lighter path for when you do not need the full machine.
AI Research on NVIDIA DGX Spark
Real notes from doing AI research on one desktop. The NVIDIA DGX Spark is a small machine with huge power (petascale means it runs about a quadrillion math steps a second), so you can push local AI further with no cloud needed. Every lesson is backed by code that runs.
Open-weight models tuned for one field each, all benchmarked on the Spark. The flagship is built with NVIDIA NeMo.
Patent
Patent Strategist
The NeMo-built patent reasoning as a small add-on patch for the base model. For your own custom builds.
Explore the modelSecurity
SecurityLLM
Tuned for cyber security questions, threat write-ups, and security know-how.
Explore the modelLegal
Saul 7B Instruct
Tuned for legal text and built to follow instructions on legal tasks.
Explore the modelFinance
Finance Chat
Tuned for finance and money questions in plain chat.
Explore the modelMedical
II-Medical 8B
Tuned for medical questions and clinical text.
Explore the modelSpace
Kepler
Tuned for space math, like the paths satellites fly. It shows short work and ends in one number you can check.
Explore the modelOpen tools whose patterns were proven on the Spark before they shipped.
Governed advisor
Orionfold Advisor
A local AI advisor over your own body of documents. Every answer names the exact source it came from. If your documents cannot support an answer, it refuses instead of guessing. Every check it passed is a saved receipt you can re-run.
Open the projectEval cockpit
Orionfold Arena
One screen to run, compare, score, and now train the AI models on your own desktop. Watch live speed and memory, rank models on a private leaderboard, queue tests and training runs, and wake up to a morning report. Nothing you type leaves your machine.
Open the projectMemory layer
Orionfold Cortex
A second brain that lives on your own desktop. It indexes your notes, stamps where every fact came from, and grades its own memory: a rebuild that would make recall worse is caught, not shipped. Your documents never leave your machine.
Open the projectOpen toolbox
fieldkit
A Python toolbox of patterns we proved on a small AI desktop. It covers the whole job: faster replies, search over your own files, scoring, training, and shipping models. Use just the parts you need.
Open the projectShort build-log posts from the real work on the machine.
Field note
My first model on a desktop
I ran my first model on a small computer on my desk. 52 milliseconds to the first word, no cloud, no per-use bill. It felt like a local function, not a service.
Read the noteField note
Access first, models second
On day one with my desktop AI machine I did not pick a model. I set up how I reach it. Models change every six months. Good access lasts for years.
Read the noteField note
The cockpit for my models
I had a shelf of models on one desktop and no way to drive them. In fifteen hours I built a cockpit to run, compare, and score them, all on that desk.
Read the noteUsed in the open
Live counts from HuggingFace, refreshed when the site builds. Built and maintained in the open by Orionfold.
- 2.4k
- Model downloads · last 30 days
- 673
- Patent Strategist · last 30 days
Start with the book. It is free to read, and every lesson runs.
Read the field log online, keep a copy if you want one, then run the open models and tools on your own machine. Want to move the work forward faster? You can sponsor it.
© 2025 NVIDIA, the NVIDIA logo, NVIDIA DGX Spark, NVIDIA Inception, NVIDIA NeMo, NVIDIA NIM, and TensorRT-LLM are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.