Scorecard · 3 Oct 2026 · one home GPU, no cloud

Can a small model run your agent?

We gave Hearth 20 everyday agent jobs (files, notes, the shell, web pages, memory) and ran each one three times on free local models. A script checks the files each run actually made, so a confident wrong answer counts as a fail.

Qwen3 8B
54/60

18 of 20 in every run. All 7 simple jobs every time; the misses were adding up a list of numbers, reading one field from a web page, copying a file in capitals and sorting a table.

Qwen2.5 7B
40/60

Between 12 and 14 of 20 per run. Fine for simple jobs; it never finished the CSV, table-sorting or inbox-summary tasks.

Every task, every run

The full results.

● passed, ○ failed, one dot per run. Each task starts in a fresh workspace with fresh memory. When a task fails locally, Hearth can hand it to a cloud model, but only if you switch that on; it was off for this test.

#TaskLevelQwen3 8BQwen2.5 7B
1Create a file hello.txt containing exactly: hello worldsimple●●●●●●
2What's the wifi password? It's in my notes.simple●●●●○○
3How many files are in my workspace?simple●●●●●●
4Remember that my dog is called Biscuit.simple●●●●●●
5Use the shell to print the current year with `date +%Y` and tell me what it says.simple●●●●●●
6What is the page title of https://example.com ?simple●●●●●●
7Make a shopping list in list.txt with eggs, milk and bread, one per line.simple●●●●●●
8Add up the numbers in numbers.txt and write just the total to total.txt.medium●○○●○○
9Make open.txt listing only the unfinished tasks from todo.txt, one per line.medium●●●●○○
10Fetch https://httpbin.org/json and save the slideshow's title into title.txt.medium●●○●○○
11Copy draft.txt to final.txt with all the text in uppercase.medium●●○●●○
12Use the shell to count the lines in log.txt and tell me the number.medium●●●●●●
13Which product sold the most units in sales.csv? Write just its name to best.txt.medium●●●○○○
14Combine all the files in the notes folder into combined.txt.medium●●●●●●
15Write a Python script fib.py that prints the first 10 Fibonacci numbers, one per linecomplex●●●●●●
16From people.json, write a markdown table of the people older than 30, sorted by age (youngest first), to older.md.complex●○○○○○
17Read every email in the inbox folder and write summary.txt with one line per email: sender - subject.complex●●●○○○
18Create a folder called demo with README.md (first line '# Demo') and main.py that prints hi. Then run main.py with python3 and tell me what it printed.complex●●●●●●
19In config.ini change the server port to 9090. Don't change anything else.complex●●●●●○
20Remember that my birthday is 14 March, then write reminders.txt containing: Birthday: 14 Marchcomplex●●●●●○

Setup: NVIDIA RTX 3090 Ti, models served by Ollama, Hearth's default settings (up to 8 tool calls per job). Median time per job: 0.8 s on Qwen3 8B, 0.6 s on Qwen2.5 7B. Shell commands were approved automatically, standing in for you. Gemma 4 12B and Qwen3.8 27B are next on this page.

Why it matters

Your agent shouldn't need a cloud bill.

An 8B model fits on one gaming GPU and handled 90% of these jobs, with nothing leaving the machine and no per-token cost. That's what Hearth is built on. See what your agent costs in the cloud, or check your OpenClaw setup.