GLM 5.3 (full) EXL3, TP6 and TP4
Serving the full GLM 5.3 model with vLLM across six or four DGX Spark nodes.
- TP6 prefill 1000 tok/s at 8K
- TP6 prose decode 27.7 tok/s (shipped, MTP k=4)
- 360K max context
Coeur d'Alene, Idaho · Serving the Inland Northwest
AI systems that do real work inside your business
We put AI on hardware you own, wire it into the software you already run, and automate the work your team keeps redoing by hand. Everything we build has to go into daily use. That is the only test that counts.
The call is free and takes about 20 minutes. If you already know your problem and want to start working, book the paid hour instead.
- Private AI on machines you own. Your data never leaves your building.
- Replace a stack of per seat subscriptions with a system you control.
- Hardware, servers, networking, and security handled by senior people.
Open recipes
Our local AI recipes
We run large models on our own NVIDIA DGX Spark cluster and publish what works: launch scripts, overlays, measured numbers, and the caveats. Free to use, on GitHub and Hugging Face.
Qwen3.8 Flash Next (NVFP4), TP2
A vLLM serving recipe across two DGX Spark boxes over their ConnectX-7 RoCE link. No image rebuild.
- ~101-103 tok/s on code, one request at a time
- ~63-66 tok/s on prose
- ~3,270-3,290 tok/s cold prefill at 32K
GLM 5.3 EXL3 GPTQ dense-decode codes
The per-rank int4 GPTQ codes used by the GLM 5.3 TP6 and TP4 builds for the dense decode path.
- 312 dense linears per rank, int4, group size 32
- KL gate: TP6 PASS (0.0284), TP4 PASS (0.0268)
- Pairs with the GLM 5.3 GitHub recipe
Numbers are single-request measurements on our cluster; each README has the conditions and caveats. See everything on GitHub →
Start with an hour
$500 / hour
Bring your questions, leave with a plan. A working session with the person who would build the system, not a sales call.
- What AI should and should not touch in your business
- Local, cloud, or hybrid, and what each one costs you
- A written summary of what we covered and what to do next
Backed by Adapt Digital Solutions
Coeur d'Alene, Idaho
Built for companies from $1M to $50M
Senior IT and cybersecurity bench
What we build
Three ways we put AI to work
Most companies do not need an AI strategy. They need one system that works, and then another one. Start wherever the pain is.
Private AI Systems
An AI machine that lives in your office. We deploy NVIDIA class local AI hardware, connect it to your documents and systems, and give your whole team access. Nothing goes to an outside model.
- No per seat fees, because there are no seats
- Your files, contracts, and customer data stay on site
- Answers limited to what each person is allowed to see
Cloud AI Systems
Custom software that connects your business to the big cloud models without handing over the keys. Your IT people can see exactly what leaves the building and what comes back.
- Connected to your CRM, files, and internal databases
- You choose what data can leave and what cannot
- One interface instead of eleven browser tabs
Workflow Automations
The quotes, intake forms, invoices, reports, and follow ups that eat your week. We find the repeat work, automate it, and show you the hours it gives back.
- We time the process before we touch it, then time it again
- A person still approves anything that matters
- Documented and handed over, so you are not stuck with us
The subscription problem
You are renting your AI stack
Every AI tool your team signs up for bills by the seat, every month, with no end date. Stack four of them and you are paying rent on software that also keeps a copy of your work.
Buy the machine once instead. Everyone in the building uses it. Your bill does not move when you hire five more people, and the data sits somewhere you can point at.
We are not against cloud AI. Half of what we build runs on it, because for some jobs the frontier models are worth every penny. But four overlapping per seat tools nobody chose on purpose is not a strategy. It is a leak.
What a private AI system replacesChat assistant seats, 40 at $30/mo$14,400
Meeting notes tool, 40 at $20/mo$9,600
Document search tool, 40 at $15/mo$7,200
Transcription and add ons$3,000
Per year, every year$34,200
Illustrative math for a 40 person company at typical per seat prices. Your numbers will be different. The shape of the problem usually is not.
Why us
Who is actually behind this
Most of the AI shops in this market opened in the last two years. Here is what sits behind this one.
Adapt Digital Solutions
This is the AI division of an established Coeur d'Alene firm that has been building, hosting, and supporting business systems for years. Same company, same phone number, same people to yell at.
Senior IT and security bench
AI lands on your network and your servers whether anyone planned for it or not. We bring in senior IT and cybersecurity people and vendor partners to handle that side instead of calling it somebody else's problem.
Nobody pays us for a pilot
A proof of concept that nobody opens after week two is wasted money. What we build goes into daily operation, with training, documentation, and a person on the phone when it breaks.
You own it
The hardware, the configuration, the documentation, the credentials. All yours. Fire us and the system keeps running, which is exactly how we set it up.
Proof
Chuck has been running his own company on this for years
Chuck has been teaching business owners how to use AI on his YouTube channel, Chuck the Contractor, since 2023. The audience there is contractors, so the examples are job sites and estimates. The systems underneath are the ones we build here.
More of it is on the channel. If you want the version aimed at your industry instead of contractors, that is what the call is for.
How it works
Consult, assess, build, run
You should be using something real inside the first month. Anyone who wants a 90 day discovery phase before you see software is billing you to learn your business.
01
Consult
A free 20 minute call to see whether there is anything here worth doing. Or the paid hour, if you already know what you want to fix.
02
Assessment
We map the workflows, the data, and the risk. You get a written plan: what to build first, what it costs, what it should give back, and which parts stay in the building.
03
Build
Hardware, integrations, custom software, training. Scoped and priced before we start, and you use a rough version early instead of waiting for a reveal at the end.
04
Run
Monitoring, updates, backups, model swaps, and the next workflow when you find it. Month to month, at a number you know in advance.
Local
We can drive to your office
We are in Coeur d'Alene. When the AI runs on a machine in your building, it helps that the people who installed it can be standing next to it in twenty minutes.
Coeur d'Alene, Post Falls, Hayden, Rathdrum, Sandpoint, Spokane, Spokane Valley, Liberty Lake: we work those in person. Everywhere else we work remotely, and we will tell you up front which parts of a job actually need somebody on site. Most of it does not.
This market is moving. The University of Idaho is starting AI degrees in Coeur d'Alene this fall, manufacturers around Spokane are already running machine vision on the floor, and there is a technology summit at the resort in October. The companies that sort their systems out this year are the ones quoting faster than their competition next year.
Who we work with
- Companies from roughly $1M to $50M in revenue
- Owner led, with someone who can make a decision
- Manufacturing, construction, professional services, healthcare administration, distribution, finance, legal
- Anyone with a good reason not to send their documents to a public AI service
Talk to us
Tell us what is eating your week
We answer within one business day. If this is not something we should be doing for you, we will say so on the first call and point you at whoever should.
- Free 20 minute discovery call, no charge and no obligation
- Paid strategy consult at $500 an hour when you want to move faster
- You talk to the people who would build it
Reach us directly
(208) 648-3977
212 S 11th St Unit #4A
Coeur d'Alene, ID 83814
Start a conversation
First name, last name, email, and phone are required. We do not share your information.