Finding
Ternary Bonsai 2 27B is Qwen3.8 27B compressed, at a sixth of the price. I ran it four times and it never rendered once. Three runs ran out of clock while still planning, and the fourth stopped early, wrote its file and never opened it. Qwen3.8 27B did look at its own render three times, but spent all three rounds on spouts.
The prompt, v1, frozen
Draw an exploded assembly diagram of a French press, as an SVG.
Bonsai wrote its file and never looked at it, and Qwen looked three times but only fixed the spouts, so neither of them noticed their filter wouldn't work.
| check · how it is decided | Qwen3.8 27B |
|---|---|
| All seven partsby eye · Pass if all seven are drawn as separate shapes and each has a label naming it. A part that is only named in a label or a comment, with nothing drawn, fails it. | ✗ fail |
| Stack ordermeasured · Knob, lid, spiral plate, mesh, cross-plate and carafe, top to bottom along the axis, read from each part's y coordinate in the SVG. Its comment is ignored. The rod is left out, because it runs through the lid. | · n/d |
| Filter fits the glassmeasured · Every filter piece drawn is 80% to 105% of the width of the carafe's glass wall, measured in the SVG. | ✓ pass |
| One assembly axisby eye · Pass if one vertical line runs through the middle of every part, and every leader line leaves from the same side. One part pushed off to the side fails it. | ✓ pass |
| Labels matchby eye · Pass if following each leader line from its label ends on the part that label names. One label pointing at the wrong part, or at nothing, fails it. | ✓ pass |
Ground truth: The standard three-piece plunger, from the top down: spiral (spring) plate, mesh screen, cross-plate. The cross-plate is the bottom piece, because its threaded stem is what the rod screws into. The spiral plate above the mesh is what presses the mesh edge out against the glass. Checked against the WebstaurantStore anatomy diagram, which is drawn exploded and labelled, and against each part's y coordinate in the SVG rather than its comment. Two-piece travel presses exist and stack differently, which is why v2 of the prompt names the design. Every model on this task.
order is left blank. That check reads three plates against each other, and
this drawing has only one of them, so failing it would count the missing parts
twice.
What's jagged here
Five models have done this task now, and four of them drew a filter that would not hold grounds back. DeepSeek 4.1 Flash got it right for four cents, and that is the number Bonsai was meant to beat.
So did compression take the knowledge? Bonsai's reasoning names the parts and argues about the gaps between them, so the knowledge is still there. What went is the step from planning to looking. It plans one geometry, writes another and stops.
Verdict
Verdict
Qwen3.8 27B draws a good-looking French press with a mesh that nothing holds in place. Bonsai drew one too, at a sixth of the price, and never found out what it looked like.
n = 1 run of Qwen3.8 27B, and 4 of Bonsai with no file. No test.
The other four presses are in the gallery.