whitepaper

A Small Shop, Not a Black Box: The Case for Private AI on Hardware You Can Drive To

August 23, 2026 · WeldonPC.ai

Small businesses adopted generative AI faster than almost any workplace technology in memory. The U.S. Chamber of Commerce's latest Empowering Small Business report found that 58% of small businesses now use generative AI, up from 40% in 2024 and more than double the 2023 rate. In a May 2025 survey of 947 small businesses backed by PayPal, 82% said adopting AI is essential to staying competitive.

The same surveys carry a second number that gets less airtime. In the Upwork Research Institute's Q1 2026 study of small and mid-size businesses, data privacy and security topped the list of adoption barriers at 49%. The PayPal-backed survey found 38% of small businesses worry about data privacy and security in connection with AI. Verizon's 2025 State of Small Business survey found that among non-users, security concerns are among the most common barriers holding them back.

In short: small firms want the tool and don't trust the pipe.

This paper describes WeldonPC's answer. WeldonPC.ai is a private AI membership from WeldonPC, the Cleveland computer-services company. Real open-weight AI models run on a GPU server WeldonPC owns, in Cleveland. No third-party cloud sits in the path between your question and the answer — member conversations never touch OpenAI, Google, or any of the giant cloud companies the industry calls hyperscalers. A small shop, not a black box.

It is also a paper that will spend a whole section telling you what our models are not good for, and when you should use a frontier cloud model instead — the industry's word for the newest, biggest models from companies like OpenAI, Google, and Anthropic. That is not modesty for its own sake. It is the only way a claim like "your conversations stay on WeldonPC hardware" stays believable.

Where the words go

Start with what the big consumer AI products say about themselves, in their own documentation.

OpenAI's help center states that ChatGPT "improves by further training on the conversations people have with it, unless you opt out." That is the consumer default: your conversations are training material until you find the setting that says otherwise.

Google's privacy hub for the consumer Gemini apps discloses that human reviewers read some collected chats, and includes this sentence, verbatim: "Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services, including machine-learning technologies." That is Google's own advice about Google's own product.

To be fair — and this paper intends to be fair throughout — the paid business tiers are different. OpenAI states that it does not train on inputs or outputs from ChatGPT Business, Enterprise, or the API by default. Google Workspace states that customer content is not used to train models outside your domain without permission. If your whole team is on properly configured business accounts, the training problem is largely a solved one — on paper.

Three things should still bother a small-business owner.

The protective defaults live on the business tiers. The consumer products most people actually type into are the ones that train by default, and the difference between the two is a settings page most people have never seen.
Even with training off, your words live in someone else's building, under a policy that company can revise, in a legal environment that keeps shifting — at least 19 states have passed comprehensive privacy laws, per IAPP's most recent report, a patchwork the U.S. Chamber itself flags as a competitiveness problem.
You are trusting a document, not an arrangement you can inspect. Nobody at a hyperscaler will take your call about where, physically, your conversation went.

That last point is the one WeldonPC.ai answers.

What WeldonPC.ai is, in plain English

WeldonPC.ai is a membership. You apply, a human reads the application, and if you're admitted you get a login to a chat application — persistent conversations the model can remember within, plus a rating button and an exact token meter on every reply, so you always know what an answer cost. (A token is a chunk of text — roughly three-quarters of an English word.)

Behind that login is the part that matters. When you type a question, it travels from your browser to a GPU server — the graphics-processor hardware AI models run on — that WeldonPC owns, in Cleveland. The model on that server reads your words and writes the answer. That moment — model reads, model writes — is called inference, and in our setup nobody else is in the path for it. Not OpenAI. Not Google. Not any cloud provider. One company, one city, hardware with our name on the invoice.

The models are open-weight models. Plain English: companies like Meta, Google, Alibaba, Microsoft, Mistral, and DeepSeek publish the actual model files — the "weights" — so that anyone with suitable hardware can download and run them, the way you'd run any other program you own. The model file itself phones nobody. We run thirteen of them, and the next section walks through every one.

One more thing, because honesty is the entire brand: the demo bot on our homepage is a scripted simulation, and it says so, in its own words, to everyone who talks to it. We built a fake bot that admits it's fake. A real model would burn GPU time on drive-by traffic; a fake bot that pretends to be real is exactly the kind of thing this paper exists to be against. Members talk to the real models.

Thirteen models, sorted by the work you do

A note before the tour. The number in each model's name — 3B, 14B, 33B — is its size in billions of parameters: the dials the model tuned during training. Bigger generally means smarter, slower, and costlier to run. Nearly every benchmark number below is vendor-reported — the score a model's own maker published on its official model card. Benchmarks are standardized tests for models, scored out of 100 unless we say otherwise. We cite them because they're the primary record, and we label them because vendors grade their own homework. Licenses are named exactly, because "open" covers everything from genuinely unrestricted to open-with-homework.

If you mostly write: drafts, rewrites, summaries, everyday questions

Llama 3.2 (3B) — free taste. Meta released this small model in September 2024 under the Llama 3.2 Community License, scoping it to summarization, rewriting, and writing-assistant work — Meta itself pitches this size class as small enough for phones and laptops. It's the lightest model in the library and the first one new accounts meet: a restricted, real-inference taste of the service before a paid tier is chosen.
Llama 3.1 (8B) — Core. Released by Meta in July 2024 with a 128,000-token context window (plain English: it can hold a few hundred pages of text in mind at once) and eight officially supported languages. Meta's own scores show its character — strong instruction-following (IFEval 80.4) and respectable math for its size, with a general-knowledge score (MMLU 69.4) that honestly signals it is not an encyclopedia. Its license is the Llama 3.1 Community License Agreement — free for any business your size or ours, but a custom license with conditions: a "Built with Llama" attribution requirement, naming rules for derivatives, and a user-count ceiling that only troubles the largest companies on earth. Open-with-homework, not open source in the strict sense.
Mistral (7B) — Core. The French lab Mistral AI released this in September 2023 under the Apache License, Version 2.0 — a true open-source license with no branding or user-count strings attached. It's a quick, capable generalist for drafting and rewriting. Two honest notes from the record: Mistral's own paper concedes that factual-recall depth is the known limit of the 7B size, and its famous benchmark wins were against 2023-era competitors, so treat them as history, not a current ranking.
Gemma 4 (9B) — Plus. The newest family in the library. Google released the Gemma 4 family in April 2026 under Apache 2.0 — a notable change, since Gemma 3 shipped under Google's custom terms — and positions the family for complex logic, code generation, structured output, and fluency in over 140 languages. It's the modern all-rounder of the mid-weights, with newer research lineage than anything else its size in our library — and one placement note worth flagging now: it is the single exception to the size-to-tier pattern, a 9B that lives in Plus rather than Core.

If you need it to show its thinking: analysis, math, working through a problem

DeepSeek R1 (8B) — Core, and DeepSeek R1 (14B) — Plus. Their full official names end in Distill, and that word deserves a plain-English sentence, because these are not shrunken versions of DeepSeek's giant R1 itself. R1 — a 671-billion-parameter model — wrote out roughly 800,000 worked examples with its reasoning shown, and smaller existing models were then trained on those examples — a student studying a master's worked solutions. Both show their reasoning in a visible thinking section before the final answer — which is half the value: you can watch it work. The 8B (built on Llama 3.1, MIT-licensed weights over the Llama base license) posts vendor-reported math scores far above its weight — 89.1 on MATH-500. The 14B (built on the Apache-licensed Qwen2.5-14B, MIT over Apache) steps up meaningfully on both math (93.9 MATH-500) and code, making it the stronger reasoner of the pair.
Phi-4 (14B) — Plus. Announced by Microsoft Research in December 2024, with its open, MIT-licensed weights following on Hugging Face in January 2025, Phi-4 is a specialist: excellent vendor-reported scores in reasoning (MMLU 84.8) and code (HumanEval 82.6) packed into 14B parameters. Its own model card is refreshingly blunt about the trade: primarily English, a short 16,000-token context window, and a factual-recall score so low (SimpleQA 3.0) that Microsoft effectively tells you not to use it as a knowledge base. Use it to reason about material you give it, not to remember things for you. Microsoft's technical report also admits it struggles with strict formatting instructions and tends to run long.
Qwen 3 (14B) — Plus. From Alibaba's Qwen team, released April 2025 under Apache 2.0. Its party trick is a hybrid design: one model that can switch between a thinking mode for hard problems and a quick mode for ordinary chat, so you're not paying thinking-time prices for "rewrite this email." Broad language support — the family claims over 100 languages.

Since two DeepSeek models sit in this library, the elephant gets named. DeepSeek's hosted app — the one that made headlines — is a legitimate concern, on three documented counts.

The company's own privacy policy says hosted-service data is collected, processed, and stored in the People's Republic of China.
Researchers at Wiz found an exposed DeepSeek cloud database leaking over a million log entries, including chat history, in January 2025.
Texas banned the app on state devices days later.

None of that applies to the open model files running on our hardware — a downloaded weight file transmits nothing to DeepSeek. One caution deserves passing along anyway, because nobody else in this business seems to: NIST's evaluation found DeepSeek models markedly more susceptible to jailbreak prompts than U.S. reference models. NIST tested DeepSeek's full-size models, not the distills in our library — which are, remember, Llama and Qwen models retrained on R1's worked examples — but because the distills are trained to imitate R1, we treat the finding as a caution for them too. For a member asking math and business questions this is a non-issue; for anyone building customer-facing automation on them, it's a real design consideration, and we'd rather you hear it from us.

If you code: websites, scripts, formulas, fixes

Qwen Coder (14B) — Plus. Qwen2.5-Coder, released November 2024 under Apache 2.0 and covering 40-plus programming languages, was trained on 5.5 trillion tokens heavy on source code. Its native context window is about 33,000 tokens — dozens of pages of code at once — and the model card describes an optional stretch past 130,000. Honest footnote: the family's headline "comparable to GPT-4o" benchmark claims belong to its 32B flagship, not this 14B. This one is simply a very good everyday coding hand.
DeepSeek Coder (33B) — Max. The old master. Released November 2023, trained on two trillion tokens that were 87% code, fluent in 87 programming languages, with a vendor-reported HumanEval-Python of 79.3 for the instruct model and a paper claiming it outperformed GPT-3.5-era closed models on code. Two honest caveats: its 16,000-token context is small by current standards, and its weights ship under the custom DeepSeek License Agreement with use restrictions — not MIT, whatever the internet tells you. It remains the deepest pure-code model in the library.

The deep end: the Max library

GPT-OSS (20B). OpenAI's open-weight model, released under Apache 2.0 on August 5, 2025 — yes, OpenAI publishes open-weight models too. It's a mixture-of-experts model: 21 billion total parameters, only 3.6 billion of which activate for any given word — think of a firm that keeps a large staff but routes each question to the few specialists it needs, so answers come faster than the headcount suggests. It offers three selectable reasoning-effort levels, from quick to deep, with a context window over 130,000 tokens.
Gemma 3 (27B). Google's heavyweight, released in March 2025, with a 128,000-token input context, image-input support, training across 140-plus languages, and Google's own intended-use list naming summarization, email drafts, and marketing copy. It ships under the custom Gemma Terms of Use, not Apache. Google's card also concedes it may generate incorrect or outdated factual statements and can miss sarcasm and nuance — limitations we'll return to, because they're true of every model in this paper.
Qwen 3 (30B). Another mixture-of-experts: 30.5 billion parameters total, about 3.3 billion active per word, Apache 2.0, with the same thinking-mode/quick-mode switch as its 14B sibling. Because the design activates only a few of its specialists per word, Qwen advertises near-small-model generation speed from a 30B-class brain. Qwen's release blog claims it outperforms their own dense 32B reasoning model while activating a tenth of the parameters; vendor-reported, as always.

Speed, measured on our own hardware

Every number in this section was measured on WeldonPC hardware, on the models members actually use. A reminder of the conversion: a token is roughly three-quarters of an English word.

The 7-9B library generates at over 100 tokens per second and answers in well under a second once warm. ("Warm" means the model is loaded and recently used.) That's faster than you can read the reply aloud.
GPT-OSS 20B measured about 74 tokens per second — a deep-end model that still moves like a small one, which is the mixture-of-experts design doing its job.
The largest models, the 27-33B class, deliberately trade speed for depth: a long, careful answer can take a minute or more. (The mixture-of-experts Qwen 3 30B writes each individual word faster by design — what takes the time is that a thinking-mode answer writes a great many of them.)

We will not dress that last number up. If you want a model that thinks at 33B depth and answers instantly, that machine exists — it lives in a hyperscale data center, not on WeldonPC hardware, and your conversation goes with it. The trade is the product.

The bill, explained like a utility bill

Most AI pricing is a flat monthly rate hiding a usage meter you can't see. We went the other way: the meter is the pricing, and you can read it on every reply.

The unit is the weighted token. Each reply's real token count is multiplied by a factor that reflects the compute a model class consumes — bigger models simply do more work per word: 0.5x for the lightest models below the 7B class, 1x for the 7-9B class, 2x for the 14B class, 4x for 20B and up. Exactly like kilowatt-hours: the meter measures what you actually drew, and the rate depends on what you drew it for.

Core — $15/month. The 7-9B library and 1 million weighted tokens a month. (The one exception noted earlier: Gemma 4, the 9B, lives in Plus.)
Plus — $25/month. Adds the 14B library; 2.5 million weighted tokens.
Max — $40/month. Adds the 20-33B deep library; 5 million weighted tokens.
Token top-ups are $5 per 500,000 weighted tokens. A single-model unlock is $5/month. Cancel anytime.
New accounts start with a free, restricted taste of Llama 3.2 3B — real inference on the real hardware, before any tier is chosen.

For context, the going per-seat rates elsewhere: ChatGPT Plus is $20/month; ChatGPT Business is $30 per user monthly, or $25 billed annually, with a two-seat minimum; Microsoft 365 Copilot runs $30 per user per month billed annually, as an add-on to a qualifying Microsoft 365 plan, with a newer small-business version promoted at $18; Google sells AI Pro at $19.99/month and folds Gemini features into Workspace plans from $7 to $22 per user.

Here is the comparison stated honestly, because a dishonest version of it would be easy to write: those subscriptions buy access to frontier models that are, on raw capability, smarter than anything in our library. What they do not buy is a machine you can drive to, a meter you can read, and a company of known size answering a phone number you already have. You are not choosing between a cheap thing and an expensive thing. You are choosing what the money is for.

The privacy promise, including the part most vendors mumble

Here is the promise, complete, in the same words we publish everywhere:

Conversations stay on WeldonPC hardware.
They are never sold, and never used to train models — ours or anyone's.
WeldonPC staff can access conversations when there is a reason to — support, billing disputes, abuse handling, lawful requests — and every access is recorded.

Notice the third bullet. Every hosted service on earth has staff who can technically reach your data; most bury that fact in paragraph forty of a policy. We print it in the promise itself, with the audit trail attached, because a privacy pitch that pretends no human can ever see anything is either lying or describing a service nobody can support. A small shop, not a black box.

Why this matters more for some readers than others:

Medical and dental offices. HIPAA's covered entities include doctors, dentists, clinics, and other providers who transmit standard electronic transactions, and HHS guidance is blunt that a cloud vendor which touches patient information on your behalf is a business associate requiring a written agreement — even if it only stores encrypted data it cannot read. We will say plainly what a compliance-minded reader should already suspect: a membership is not a compliance certificate, and that rule applies to any vendor handling patient data — including us. If your practice handles PHI, talk to your compliance advisor, and to us, before typing patient information into anything. The difference a local operator makes is that the where-does-the-data-live question has a one-sentence answer, and the company holding it takes your call.
Law offices. The ABA's first ethics opinion on generative AI, Formal Opinion 512, warns that self-learning AI tools "by their very nature, raise the risk that information relating to one client's representation may be disclosed improperly," and requires informed client consent — not boilerplate — before client information goes into such a tool. Our models never train on conversations, which removes the self-learning mechanism the opinion worries about; the duties of competence and verification stay with the lawyer, exactly where the ABA put them.
Tax preparers, auto dealers, financial advisors. The FTC's Safeguards Rule treats tax preparation firms, financial advisors, and most auto dealers who finance vehicles as financial institutions, obligated to run a formal information-security program — nine required elements, starting with a designated Qualified Individual — and, since May 2024, to report qualifying breaches to the FTC within 30 days. Every new place customer information flows is a line item in that program. A single, known, local destination is an easy line to write.

What these models will not do

This section is the reason to trust the rest of the paper.

They are not encyclopedias. Small models store less of the world. Microsoft's own card scores Phi-4 at 3.0 on the SimpleQA factual-recall benchmark against GPT-4o's 39.4. OpenAI's own model card reports GPT-OSS 20B hallucinating on 91% of SimpleQA factual questions, and says plainly this is expected of smaller models. Ask our models to work on text you give them — summarize, draft, extract, reason — and they shine. Ask them obscure trivia cold, and they will sometimes answer confidently and wrongly.
They do not know current events. Llama 3.1 and 3.2's training knowledge stops in December 2023; GPT-OSS's in June 2024. Anything after a model's cutoff has to be pasted into the conversation; the model cannot look it up.
Even summaries need a human eye. Vectara's hallucination leaderboard, which tests models on summarizing only a supplied passage, still measures small open models inventing content in a few percent of summaries. The house rule we recommend every member adopt: anything that feeds a decision gets a human read first.
Long autonomous work is out of scope. METR's research found that even the frontier models of early 2025 had a time horizon of only about an hour — meaning a 50-50 shot at completing a task that takes a human that long, with near-reliable performance only on tasks of a few minutes. Small local models sit far down that curve. Our models are excellent assistants for a task you are driving. They are not employees.
They are not your lawyer, doctor, or accountant. ABA Opinion 512 puts verification duties on the professional, not the tool, and that is the right model for every regulated field: the machine drafts and triages, the licensed human decides.

And the sentence a sales page would never print: sometimes the right tool is a frontier cloud model. Genuinely novel, hard problems; research that needs current knowledge of the world; sprawling multi-step work across a big codebase; the highest-stakes drafting of your year — for those, a business-tier frontier subscription with training off by default is a defensible choice, made with open eyes. Plenty of businesses will sensibly run both: the private machine for the everyday work that involves the business's actual information, the frontier cloud for the occasional problem that needs a bigger brain and contains nothing sensitive. We would rather you get the right answer somewhere else than a wrong answer here.

How to get in

Membership is limited — the hardware is real, so the roster is finite. Admission is by application: a $5 fee, personally reviewed, and credited toward your first month if you're accepted. No questionnaire-shaped robot decides; a person reads it. If you want a feel for how the service talks before applying, the homepage demo is open to everyone — it's the scripted stand-in, and it will tell you so itself. The real models start on the other side of the application.

Cancel anytime. The meter, the promise, and the phone number don't change.

About WeldonPC

WeldonPC is a computer-services company in Cleveland, Ohio, owned by Weldon Hastings, serving Greater Cleveland with computer repair, websites, managed IT, electronics recycling, and data recovery. WeldonPC.ai runs on hardware WeldonPC owns. Reach us at (216) 475-6000 or weldonpc.com.

Sources

← All news