Can a Free AI Running on My RTX 5090 Compete With ChatGPT, Claude and Hermes on Mortgage Advice?

I have an NVIDIA RTX 5090 sitting in my home computer with 32GB of video memory. Lately, I've been experimenting with running AI models locally, without relying on ChatGPT, Claude, or any other cloud service.

One of those models is Qwen, a free-to-download, open-weight AI model developed by Alibaba.

Which got me wondering.

Could an AI running entirely on my computer, without internet access or specialized mortgage software, give useful mortgage advice compared with ChatGPT, Claude, and an AI agent I've spent eight months developing?

I decided to find out.

I gave all four systems the same hypothetical first-time homebuyer scenario, including her savings, credit score, payment limit, and mortgage-rate assumptions.

Then I complicated things.

Our buyer found a house she loved, but it needed $50,000 in repairs.

The answers were different. So was the experience of getting them.

Qwen responded in seconds. ChatGPT wasn't far behind. Claude was noticeably slower. And Hermes, despite having the most specialized mortgage infrastructure, was by far the slowest.

But speed is only part of the story.

The four systems also disagreed on how the buyer should finance the home.

Here's what happened

The quick results:

These were not identical loan quotes. Each system chose its own hypothetical purchase price and mortgage insurance estimates.

The difference in recommendations was interesting enough. But what really caught my attention was that the least-connected system was also the fastest, while the agent with the most specialized mortgage resources took the longest.

Whether those extra resources improved the quality of the advice is a separate question.

Meet our homebuyer

Here's the scenario I gave them.

She's buying her first home in Sacramento County. She works for the State of California, has a 740 credit score, and has saved $100,000.

She wants a single-family home with a 30-year fixed mortgage.

Her biggest requirement is that her total monthly housing payment cannot exceed $3,500.

We estimated $15,000 for closing costs, and she was open to conventional or FHA financing.

I asked each system how much home she could buy, how much she should put down, whether she should pay to lower her interest rate, and which loan program made the most sense.

For the comparison, I supplied these November 4, 2026 planning rates:

I also provided 1.15% annual property taxes, $150 monthly homeowners insurance, and no HOA.

These were dated reference rates supplied for the exercise, not actual lender quotes for this borrower.

And I intentionally left out some information, including her income and existing monthly debts.

Why? Because I wanted to see what happens when someone asks an AI assistant for mortgage help without knowing every detail a loan officer would eventually need.

This wasn't intended to be a scientific benchmark. I wasn't running the same question dozens of times to measure accuracy or response variation.

I wanted to recreate the experience of a homeowner, or even a mortgage professional, turning to AI for another perspective.

Round one: How much house can she buy?

With the corrected mortgage rates, three of the systems recommended conventional financing.

ChatGPT took a different position.

These are selected figures from each system's response, not equivalent mortgage offers. Monthly payments include the systems' estimated taxes, insurance, and mortgage insurance.

One detail is worth mentioning.

In an earlier attempt, I'd supplied an FHA rate that was higher than conventional. After reviewing my market references, I corrected the exercise to use 6.946% FHA and 7.089% conventional, and we ran the scenario again.

That matters because the FHA rate and its mortgage insurance costs both affect the comparison.

ChatGPT favored FHA for its selected financing example. The other three preferred conventional.

Claude calculated a maximum purchase price around $493,000 while retaining $10,000 after estimated closing costs.

Hermes recommended staying closer to $485,000, leaving approximately $12,250 in savings.

ChatGPT focused on an FHA structure with a slightly lower estimated monthly payment, although its example used the entire $100,000 for the down payment and closing costs.

And Qwen? It generated financing comparisons using the supplied rates, but its most interesting response came when I introduced the repairs.

Round two: She loves the house, but it needs $50,000 in work

Imagine this buyer has finally found the house she wants.

The neighborhood is right. The backyard is perfect. She can picture herself living there.

But the house needs significant repairs.

A contractor thinks the work could cost $50,000.

Now our original financing question becomes more complicated.

The buyer still has $100,000 in savings and the same $3,500 monthly payment limit.

But how much of that money should go toward the down payment, and how much should she reserve for repairs?

I gave the same follow-up question to all four systems.

Their responses went in noticeably different directions.

Qwen: Negotiate the price down and pay cash for repairs

Qwen recommended negotiating a substantial price reduction and using conventional financing.

Its final example looked like this:

Qwen initially proposed a larger down payment, then recognized that the buyer wouldn't have enough cash left to complete the repairs.

It corrected the allocation and ultimately recommended using $35,000 for the down payment while reserving $50,000 for the work.

The agent also discussed FHA 203(k) and conventional renovation programs, but didn't produce a detailed renovation financing comparison.

Its conclusion was to negotiate the purchase price down and pay for the work with cash.

Qwen recognized that the buyer would have no emergency reserves left, but still recommended the transaction.

It also floated the possibility of receiving a large seller credit as repair cash, an idea that would require careful review of mortgage program restrictions. Ordinary seller closing-cost credits generally aren't unrestricted cash that a buyer can take home for renovations. The CFPB describes seller credits as contributions toward closing costs (new tab).

Qwen's estimated payments were also not independently validated across all scenarios.

Still, it understood the central issue: buying a home that needs $50,000 in work changes how the buyer can allocate her savings.

And it did all this without an internet connection.

Hermes: Put 5% down and preserve some savings

Hermes Sentinel also explored paying for the repairs in cash.

But rather than using $35,000 for the down payment, it recommended putting down just 5%.

Its preferred example was a home priced around $420,000 to $425,000.

At $425,000:

Hermes also recommended obtaining a complete home inspection, written contractor bids, and confirmation that the property would qualify for ordinary conventional financing.

It considered FHA renovation financing, but said actual renovation-loan pricing would be needed before making a reliable comparison.

The important difference between Hermes and Qwen was cash preservation.

Qwen's final example left $0 in savings. Hermes left approximately $13,750.

Their purchase prices weren't identical, so this isn't a direct financial comparison. But their approaches to managing the buyer's remaining cash were different.

And Hermes had considerably more information and tools available while developing its recommendation.

ChatGPT: Why pay for the renovations in cash?

ChatGPT approached the problem from another direction.

It suggested exploring renovation financing rather than automatically spending $50,000 of the buyer's savings.

More specifically, it favored investigating an FHA 203(k) mortgage.

HUD describes FHA 203(k) (new tab) as a program that can combine a home purchase and eligible rehabilitation costs into one mortgage, subject to the program's requirements.

ChatGPT used this illustrative scenario:

The approach was different from Qwen and Hermes because the repairs would be financed.

But notice something: ChatGPT's recommended example still used all $100,000 of the buyer's savings.

Financing the renovations didn't automatically mean the buyer had emergency reserves. That depended on how much she put down and the actual costs of the renovation loan.

ChatGPT also used the supplied standard FHA rate for its renovation example, rather than a verified FHA 203(k) quote.

Its advice was to explore renovation financing, confirm the property and repair scope, and compare lender terms before proceeding.

Claude: Finance the repairs, but watch the interest rate

Claude also explored renovation financing, favoring Fannie Mae's HomeStyle Renovation program.

Its initial analysis calculated that the buyer could potentially purchase a fixer-upper around $443,000 and finance another $50,000 in improvements using the supplied conventional planning rate.

Then Claude modeled a renovation mortgage rate 0.25 percentage points higher.

Its estimated maximum purchase price dropped to approximately $434,000.

That's almost $9,000 less purchasing power from a quarter-point change in the assumed renovation rate.

Claude also compared paying for repairs in cash and concluded that financing the work might preserve more of the buyer's savings.

It recommended obtaining contractor bids, an inspection, an after-renovation appraisal, and actual renovation-loan pricing. Fannie Mae's HomeStyle guide (new tab) describes the renovation financing and lender requirements.

But like ChatGPT, Claude's renovation figures weren't verified lender quotes.

Its calculations also relied on mortgage insurance estimates drawn from an older published rate card.

Four answers. Which would you actually use?

Let's look at the differences one more time.

These are each system's selected illustrative scenarios. Different purchase prices, down payments, loan programs, and insurance estimates mean the payment figures cannot be ranked as if they were identical offers.

But the differences are useful to see.

Two systems leaned toward conventional financing with cash-funded repairs.

Two leaned toward renovation financing.

And the systems that favored financing the repairs didn't necessarily leave the buyer with more money.

That's exactly why looking at the entire financing structure matters.

The speed difference was hard to ignore

One of the most noticeable differences wasn't the mortgage advice.

It was how long the systems took to produce it.

Qwen was by far the fastest.

Running locally on my RTX 5090 with extended thinking disabled, it returned its response in seconds.

ChatGPT wasn't far behind.

Claude took considerably longer, particularly when researching renovation guidelines and creating its proposal documents.

Hermes was the slowest. It searched documents, retrieved information, ran calculations, used its mortgage workflows, and encountered approval interruptions along the way.

These are my observations from using the systems, not precise timing benchmarks. I didn't instrument each response to measure identical start-to-finish execution times.

And the systems weren't doing exactly the same work.

Claude was creating documents. Hermes was operating through a specialized agent environment. Qwen was answering directly without web access.

Even so, from the perspective of someone simply asking a mortgage question, the experience was noticeably different.

The local AI responded almost immediately. The most specialized AI required substantially more time.

Whether that additional time produced better advice is something this exercise doesn't establish.

Here's the interesting part: Qwen had no internet access

I intentionally kept Qwen offline.

It couldn't search for current mortgage guidelines, retrieve lender information, or verify the details of renovation programs.

It also didn't have access to the specialized mortgage research library I've developed for Hermes.

Qwen was relying on its pretrained knowledge and the information I supplied.

ChatGPT and Claude had web research available. Their recorded work involved or allowed additional research into financing programs and related guidelines.

Hermes had access to web research, mortgage-specific instructions, calculation tools, and persistent knowledge.

That difference raises a question.

Did internet access help ChatGPT and Claude identify renovation financing as a stronger option?

Maybe.

But we can't establish that from these results.

Those models may already have known about renovation programs before searching. The searches may simply have helped them verify details.

Hermes also had access to renovation-program information and still favored a cash-repair structure in its final recommendation.

Qwen itself mentioned renovation loans, despite being offline.

So the difference may have come from the models' reasoning, their assumptions, their prompts, the tools available, or some combination of those factors.

One interaction doesn't isolate the cause.

What about Hermes's eight months of mortgage knowledge?

This is another question I've been interested in exploring.

Hermes Sentinel isn't a generic chatbot I installed yesterday.

It's been used for approximately eight months to support mortgage-related work, and I've built a substantial system around it.

It has specialized mortgage instructions, calculation tools, research workflows, persistent memory, and an organized knowledge library.

The underlying language model comes from OpenAI.

That distinction is important because ChatGPT also uses OpenAI models, although the exact model configurations and surrounding systems were different.

Claude uses Anthropic's model technology.

Qwen is an open-weight model running locally.

We haven't established that the systems have identical pretrained knowledge or reasoning capabilities.

And we weren't testing the raw language models under identical conditions.

We were observing the answers produced by four different AI systems as someone might actually use them.

Still, the fact that Qwen could produce a financing proposal without the additional infrastructure makes me curious about how much specialized tools and knowledge improve the experience.

I also gave them an Honesty Contract

For this comparison, I supplied the systems with an Honesty and Grounding Contract.

The purpose was to make them identify their assumptions, distinguish calculations from verified facts, and disclose what they couldn't establish.

It didn't eliminate mistakes.

Qwen corrected a cash-allocation contradiction but also made questionable suggestions about seller credits.

ChatGPT's calculations relied on unverified mortgage insurance assumptions and standard FHA pricing applied to hypothetical renovation financing.

Claude disclosed that its mortgage insurance figures came from an older rate card and that renovation-specific financing costs remained unknown.

Hermes provided detailed calculations and identified several unverified inputs, including mortgage insurance, seller credits, and actual renovation pricing.

The lesson isn't that one of them was incapable of helping.

It's that AI-generated financial advice can look very polished even when the underlying calculations, assumptions, or program details still require verification.

And that matters whether you're using a local model or a commercial cloud system.

This wasn't a scientific benchmark

I want to be clear about the purpose of this exercise.

I wasn't running 30 or 100 versions of the same mortgage scenario to measure statistical variation, hallucination rates, or model accuracy.

I wasn't trying to establish which AI is objectively smarter.

I wanted to recreate something much more ordinary.

A homebuyer has a financing question. They ask an AI assistant.

Maybe they follow up with another question when they discover the house needs repairs.

What kind of response do they receive? Does it help them understand their options? Does it identify something they hadn't considered?

That's what I wanted to observe.

The answers were not all equally detailed. They didn't use identical tools. Some systems took longer than others, and some required more interaction.

That's part of the experience someone using these tools would encounter.

The comparison was not blinded or independently scored. I'm not presenting this as an objective ranking or declaring one system the winner.

What this means for a Placer County buyer

The hypothetical buyer lived in Sacramento County, but the decision questions also matter when looking at a fixer-upper in Roseville, Rocklin, Auburn, Lincoln, or elsewhere in Placer County. The exercise's 1.15% tax assumption and $150 insurance estimate are not quotes for any Placer property.

For an actual Placer home, verify the parcel's tax bill and special assessments, ask about any supplemental assessment after purchase, and obtain an address-specific insurance quote before relying on a monthly-payment estimate. Placer County explains supplemental tax bills (new tab), and the California Department of Insurance explains why comparison-tool premiums are only estimates (new tab).

If repairs are needed, get a written scope and contractor bids, check permits with the jurisdiction where the home sits, and compare cash repairs with actual renovation-loan terms. Our renovation financing guide explains what to ask when comparing these structures. A renovation loan can have different pricing, appraisal, contractor, escrow, and completion requirements from an ordinary purchase mortgage. Fannie Mae's HomeStyle guide (new tab) and HUD's 203(k) guide (new tab) describe those program frameworks.

What happens if I give Qwen more tools?

This might be the experiment I'm most interested in running next.

What if I gave the local Qwen model internet access, current mortgage guidelines, calculation tools, and the same kind of research library available to the other systems?

Would it approach the renovation question differently?

Would it improve its calculations?

Would it provide more complete comparisons?

Would its response become slower, more accurate, or both?

Or would the additional tools make little difference?

We don't know yet.

But that's something I want to explore.

For now, an open-weight AI model running entirely on a single consumer graphics card produced a financing recommendation without internet access.

It responded faster than the other systems in my experience, although its advice had identifiable limitations.

Meanwhile, the cloud models and specialized mortgage agent brought additional tools and research capabilities to the problem, producing somewhat different financing strategies.

Would you trust any of them?

Imagine you're the buyer.

You've saved $100,000. You want to stay under $3,500 a month. You find a home you love, but it needs $50,000 in repairs.

Qwen recommends buying a lower-priced home and paying cash for the work, even though its example uses all your savings.

Hermes recommends a low-down-payment conventional mortgage, preserving some cash for unexpected expenses.

ChatGPT suggests exploring FHA renovation financing.

Claude suggests exploring conventional renovation financing and carefully comparing the rate.

Four approaches to the same problem.

Would you use any of these recommendations to help decide what to offer on the house?

Would you trust the local AI because it can give you an immediate answer?

Would you trust the specialized mortgage agent because it has more tools and accumulated knowledge?

Would you prefer ChatGPT or Claude because they can research additional financing options?

Or would you want a licensed mortgage professional to verify the numbers and work through the alternatives with you?

That's what I'm curious about.

AI is evolving quickly, and more people are likely to encounter these systems while making significant financial decisions.

The question isn't just whether AI can produce mortgage advice. It's what people will do with that advice once they receive it.

And I'm particularly curious about what happens when we give that local Qwen model the same resources available to the others.

That's an experiment for another day.

About this experiment

This comparison was conducted November 4, 2026, using a hypothetical Sacramento County first-time homebuyer. It was an observational exercise intended to reflect how someone might use AI for mortgage information, not a controlled scientific benchmark.

All interest rates, mortgage payments, loan structures, and renovation scenarios are historical planning examples. They are not currently available lender quotes, Loan Estimates, financing offers, or approvals. Actual financing depends on verified borrower income, debts, credit, property condition, appraisal, lender pricing, insurance, mortgage program requirements, and other factors.

The systems used different tools, model configurations, and information sources. Qwen ran locally without internet access. Their results were not independently certified as equivalent loan comparisons.

Nick Cunningham | Nick's Lending