UNTETHEREDCHAOS
Lab notes · September 2026

QWEN

A free model that runs on a laptop. Fast, private, and sloppy.
Worth using only when something checks its work.

1 · It writes

Qwen drafts, on the laptop

Qwen3 30B, an open model. Nothing leaves the machine and nothing is billed. A 40-word list came back almost instantly.

1.7sfor 40 words
2 · Code grades it

A checker throws out the bad parts

A dictionary, a test suite, a rule scan. Of the 30 distinct words, one was not a real word; ten of the 40 were repeats.

29 / 30passed the dictionary
3 · A person looks

Someone checks what it looks like

Qwen reads text only. It cannot see a screenshot, so how a game looks and feels is judged by a person or a bigger model.

4 · It ships, or not

Only what passed every step

Nothing Qwen writes goes live on the strength of the model alone.

The checker is the product. The model is the cheap part.

What it is. Qwen3 30B is an open-weight model running through Ollama on the studio laptop. It has 30 billion parameters but uses about 3 billion for each word, so it answers at small-model speed. In this setup it can hold about 32,000 tokens at once (roughly 24,000 words), counting the instructions, the code it reads, and what it writes.

What the first test showed. Asked for 40 basketball words for a word game, it answered in 1.7 seconds, at about 127 tokens a second. The game's dictionary accepted 29 of its 30 distinct words. But a quarter of the list was repeats, and several words were only loosely on theme. Fast and free, not careful.

What that means. Use it for repetitive work that a program can grade: lists, test cases, one small file at a time against written rules. Don't hand it a whole game at once, and don't let it review code on its own. Instructions written for a large model also tend to work worse on a small one, so anything Qwen reads has to be written for Qwen.

Play it

Snake, with Qwen's code inside

The rules, the drawing and the controls running below are Qwen's code, exactly as it wrote them. The page around them is ours, because Qwen's own page never started the game. Click the board, then use the arrow keys or WASD. On a phone, swipe or use the buttons.

Watch the food. It only ever appears in two squares, taking turns. That is a bug in Qwen's rules that all 17 tests missed: it never moves its random number forward, so every new piece of food comes from the same two draws. We left it in.

The experiment · run 14 Sep 2026

Can a short style file teach it to build our kind of game?

The studio's games follow house rules: small files, a pure rules engine with tests, graphics drawn in code, no downloaded assets, sound off until you turn it on, no motion when a device asks for less. The experiment gives Qwen one short file of those rules, then asks it to build a tiny game one file at a time. Tests and a rule scan grade every file; a person grades the screenshots.

Result

Useful for logic only.

Qwen wrote the rules, the drawing and the controls, and each passed its tests with no human edits. It could not wire them into a working page in three tries. The game never started outside reduced-motion mode, and the touch buttons ignored a house rule even after the failing check was shown back to it.

The test game is Snake on a 20 by 20 grid. Its state is a handful of whole numbers, so the rules can be tested exactly: where the head is after a move, growing when it eats, dying at a wall. We picked it over Pong, whose bouncing maths only tests to a tolerance, and because Snake can stay fully playable with motion off by moving one step per key press.

Everything that grades Qwen is written before Qwen sees anything: a one-page spec, 17 tests, a browser check, and a scan for house-rule breaks. Qwen never writes its own tests, because a model that writes the code and the test together tends to test what it built, not what was asked for. Each lesson gets one try and two retries. A retry only sees the lines that failed.

LessonQwen writesIt passes whenWhat happened
1 · RulesThe game logic, with no page or timer code11 of 11 tests pass; the rule scan finds nothingPassed on try 2. Try 1 forgot to export two constants; shown only that error, it fixed it. 32 s.
2 · DrawingThe board, drawn with rectangles and paths only3 of 3 tests pass; no images or downloadsPassed first try. 21 s.
3 · ControlsArrow keys, WASD, and swipes turned into moves3 of 3 tests passPassed first try. 27 s.
4 · The pageThe game loop, sound switch, and pageBrowser check passes on phone and desktop; screenshots look rightFailed all 3 tries. Space never started the game; touch buttons used clicks instead of pointer events. 63–119 s a try.

The verdict was fixed before the run, so the bar could not move.

  1. Useful, build more with Qwen: all four lessons pass their check within the retry budget, Claude edits zero lines of Qwen output, and the screenshots at 390px and 1200px show a playable board.
  2. Useful for logic only: lesson 1 passes; a later lesson fails. Qwen writes tested logic; the page work goes elsewhere. This one
  3. Not useful: lesson 1 fails, or any lesson needs a Claude edit to pass.

Plan by Fable, reviewed before the run. Two problems in the setup, not in Qwen, were caught and fixed before any Qwen output existed, and are logged in the plan. One gap was ours: the page lesson showed Qwen the controls function's name but not what it expects, which explains one of its page bugs but not the two that failed it. Settings: Ollama 0.33.3, thinking on, temperature 0.6, seed 42. One game, one run: a data point, not a verdict on the model.

Findings log

Every run, as measured

DateTestResultStatus
13 Sep 2026 40 themed words for a word game, graded by the game's dictionary 1.7 s · 127 tok/s · 10 repeats · 29 / 30 real words Done
14 Sep 2026 Build Snake from the style file, four lessons, graded by 17 tests and a browser check 3 of 4 lessons passed · page failed 3 tries · useful for logic only Done
14 Sep 2026 Played Qwen's Snake logic inside our page Playable. Food only spawns in two squares, a rules bug the 17 tests did not catch Done
14 Sep 2026 Rerun, 3 seeds each: fix the food bug from a failing test 3 of 3 fixed it on the first try Done
14 Sep 2026 Rerun, 3 seeds: build the page again, this time told exactly what the controls function expects 1 of 3 passed the checks, and that page barely plays (tiny board, taps ignored). Page wiring moves to our GAMEBOYO console; Qwen writes cartridge logic Done
14 Sep 2026 BLOCKS, a falling-block game, 3 seeds: pieces and rules, graded by 56 tests and 20,000 random-input frames 3 of 3 passed. Checking against our correct version afterwards found a real bug the tests missed in 2 of 3 (the preview piece gets skipped) Done
14 Sep 2026 BLOCKS cartridge for GAMEBOYO, 3 seeds 0 of 3 when our prompt left out the board size and piece helpers; 3 of 3 once it included them. The seed-42 version is playable on GAMEBOYO, bugs and all Done

Numbers are from single runs on one laptop. They show what happened, not a benchmark.

The question is not whether Qwen is smart. It is whether a cheap draft plus a strict checker beats doing the work another way.
← All research