Studio 111 Labs · Local AI · Speed test

Pandora's Box: Speed & Convenience vs. Cost & Security

Pandora's Box. Speed and convenience vs. cost and security. An open box glowing red, with points of light escaping from it. $500 to $5,000: four boxes against one frontier model.

Cloud AI is fast and convenient. A box in your office costs money up front and keeps your files in the building. Drop in a financial file and watch the trade: the same analysis written at the speed of a frontier cloud model and four local machines, from $500 to $5,000.

Step 1 · Give it a file

Drop a CSV or Excel file here

A bank export, a ledger, a sales report. Anything with amounts in a column.

● Read inside this browser tab. Nothing is uploaded, to us or anyone.

01What you just watched

Every lane wrote the same words. The only thing that changed was how long you waited. For a summary like this one, the gap between the cloud and a well-chosen box is a few seconds. On the smallest box you wait about seven times as long.

Speed

Speed comes from the hardware and the size of the model. A bigger model on the same box writes slower. That trade is the first thing we size.

Accuracy

The totals in the demo were added up by code, so they match in every lane. The two percentages on each lane are something else: how that model scores on two published tests, one easy to relate to and one hard on purpose. The next section explains both.

02Two tests, two different jobs

Accuracy depends on the job. So each lane carries two published scores for its model: one for reading financial statements, and one for hard math with nothing checking the work. Neither is a score for your files or for the demo above.

Job oneReading the statements

A test from researchers at ETH Zürich asks 250 questions about 10-Q-style financial tables: look up a figure, add a few lines, find the highest quarter, work out gross profit. The correct answer is calculated from the table itself, so the answer key involves no judgment call.

Share of 250 statement questions answered correctly
$5,000 box · gpt-oss-120b
98.0%
$1,000 box · a 9B model, the size it can run
90.4%
Top score in the test · a 31B open model
98.4%

V-FiLLM, 10-Q setting, questions asked in one go. Six models were tested. No frontier cloud model was among them, and neither was a 20B or a 4B model, so those lanes read "not tested". The paper describes its 10-Q tables as real filings in one place and as synthetic statements modeled on a 10-Q in another. Bars start at zero.

One caution from the same paper. On a harder question set, the 9B model fell to 55.2% when asked everything at once. Broken into steps, it came back to 86.8%. On a small box, how you ask matters.

Job twoThe bar everyone is chasing

The second test is hard on purpose. FinanceReasoning's hard set is 238 calculations of about eight steps each, like a multi-step mortgage problem, and an answer counts only if it lands within 0.2% of the correct figure. The best cloud models set the bar at 91.60%. Everything else is distance to that line.

Share of 238 hard finance calculations answered correctly

Scores as published by AIMultiple, updated 17 Nov 2026. Each model reasoned step by step on its own; the source does not say whether tools were allowed. Models small enough for the $1,000 and $500 boxes were not in that test. Bars start at zero.

How close is the box?

The $5,000 box's model scores 81.09% today. In the test's original 2025 paper, the best cloud model scored 81.1% reasoning step by step. Roughly, a box in your office is where the frontier was when this test came out. Those are two separate test runs, so read it as a rough marker.

What the gap costs

At 81%, about one hard calculation in five comes back wrong when nothing checks it. At 68%, about one in three. That is too many to trust unaided.

What closes it

The test's authors found that models did better when they wrote a short program to do the arithmetic. One went from 65.6% to 83.6% that way. A good setup has the model explain and software calculate.

Why we still test first

A published score is a starting point. Before anyone buys a box, we run candidate models against your own documents and questions, and count what they get right.

03Four boxes, four budgets

Hardware budget only, with one example machine each. Prices move monthly, so read these as an order of magnitude.

Setup, model selection and training are quoted separately. Example hardware, speeds and accuracy scores per the sources below. Speeds are for one person asking at a time with a short prompt, and fall as documents get longer.

What this page is and is not. It is an illustration of waiting time, built from published speed figures. It is not a benchmark of your workload. The accuracy figures come from two published third-party tests, not from this page and not from your file, and "not tested" means no published score exists for that model. The speeds for the $500 lane and the frontier lane are Studio 111 Labs working assumptions, because public figures for both vary widely. The $2,000 lane's speed is our own estimate, scaled by memory bandwidth from a figure measured on a faster chip. Prices, speeds and scores are as published in October and November 2026. Time spent reading the prompt and network delay are not modeled. The sample ledger is synthetic.
Studio 111 Labs · Local AI

Want to see your own file on the real thing?

A first conversation covers what you want the model to do, which files it would need, and whether a local box, a business-tier cloud contract, or both is the right fit. If it is a box, we run your actual documents on candidate hardware before you buy anything, then handle procurement, the model and the training.

Start a conversation →

doug@studio111labs.com

Sources
ComputingForGeeks, Mac mini vs mini PC vs GPU for local LLMs, Oct 2026: Mac mini and Mac Studio prices and estimated writing speeds; the gpt-oss-120b figure it cites from Hardware Corner (measured on a MacBook Pro with the same M5 Max chip, 4K context). OpenClaw DC, best local LLM by Mac memory: a 92.36 tokens/sec result for gpt-oss-20b on an M4 Max and the bandwidth-scaling method. The $2,000 lane's 69 is our own scaling of that result to the M5 Pro's 307 GB/s (92.36 × 307 ÷ 410). Larsen et al., V-FiLLM: Verified Financial LLM Reasoning Benchmark: statement-reading scores, 10-Q setting, and the 55.2% and 86.8% figures for the 9B model. AIMultiple, benchmark of LLMs in finance, updated 17 Nov 2026: accuracy scores on the 238-question hard subset. Tang et al., FinanceReasoning: test design, the 0.2% tolerance, and the 65.6% to 83.6% result with program-written arithmetic (GPT-4o), and the 81.1% step-by-step score for OpenAI o1. $500 lane speed and frontier lane speed: Studio 111 Labs working assumptions, not measurements.
← Back to Studio 111 Labs © 2026 Studio 111 Labs · Doug Larsen