Moka Benchmark: AI Models Build and Brew a Moka Pot in 3D

One prompt per test, pasted verbatim, no hints and no fixes. Each coding model builds a moka pot in Blender and films it, then animates how the pot brews coffee in three.js. A checker script measures what can be measured, a judge scores what a viewer sees against written anchors, and every point carries its reason.

Leaderboard

Model Moka score Blender JS
DeepSeek V4.1 Flash 1017 63.7 62 66.3
Hy4 Flash 62.1 58.1 68.1
GLM 5.3 Flash 62.1 51.1 78.5
Qwen 3.8 Flash Next 49.4 31.7 75.9

How much VRAM each model needs to run at home: DeepSeek V4.1 Flash 1017, Hy4 Flash, GLM 5.3 Flash, Qwen 3.8 Flash Next.

The Moka score weights Blender 60% and JS 40%. Blender is the harder test and the better measure of spatial reasoning in a real 3D scene: the model has to build every part, fit them together at real scale, then move a camera around them, light the stage and render the film. The JS test starts every model from the same pot we supply and scores how well it animates the brew. The Moka score is shown only for a model that ran both.

Moka Blender v1.1

DeepSeek V4.1 Flash 101762/ 100
Construction
23.5 / 30
Film
10.5 / 20
Look
25 / 40
Efficiency
3 / 10
1 pass, 338K tokens, opencode Go, 1 run, 2026-11-21
Hy4 Flash58.1/ 100
Construction
19 / 30
Film
11.5 / 20
Look
19 / 40
Efficiency
8.6 / 10
1 pass, 101K tokens, opencode Go (free), 1 run, 2026-11-19
GLM 5.3 Flash51.1/ 100
Construction
15.5 / 30
Film
12.5 / 20
Look
15.5 / 40
Efficiency
7.6 / 10
1 pass, 136K tokens, opencode Go, 1 run, 2026-11-21
Qwen 3.8 Flash Next31.7/ 100
Construction
12 / 30
Film
6.5 / 20
Look
8.5 / 40
Efficiency
4.7 / 10
1 pass, 239K tokens, Alibaba official API, 1 run, 2026-11-20

Moka JS v1.1

GLM 5.3 Flash78.5/ 100
Process
26.5 / 30
Look
18 / 30
Clarity
17 / 20
Checks
9 / 10
Efficiency
8 / 10
1 pass, 120K tokens, opencode Go, 1 run, 2026-11-21
Qwen 3.8 Flash Next75.9/ 100
Process
25.5 / 30
Look
18.5 / 30
Clarity
17 / 20
Checks
9 / 10
Efficiency
5.9 / 10
1 pass, 196K tokens, Alibaba official API, 1 run, 2026-11-20
Hy4 Flash68.1/ 100
Process
20.5 / 30
Look
15.5 / 30
Clarity
14 / 20
Checks
9 / 10
Efficiency
9.1 / 10
1 pass, 83K tokens, opencode Go (free), 1 run, 2026-11-19
DeepSeek V4.1 Flash 101766.3/ 100
Process
22.5 / 30
Look
16.5 / 30
Clarity
14 / 20
Checks
9 / 10
Efficiency
4.3 / 10
1 pass, 254K tokens, opencode Go, 1 run, 2026-11-21

The films

DeepSeek V4.1 Flash 1017: Blender v1.1 62 / 100, JS v1.1 66.3 / 100
GLM 5.3 Flash: Blender v1.1 51.1 / 100, JS v1.1 78.5 / 100
Hy4 Flash: Blender v1.1 58.1 / 100, JS v1.1 68.1 / 100
Qwen 3.8 Flash Next: Blender v1.1 31.7 / 100, JS v1.1 75.9 / 100

How the test runs

  • Blender: model the pot, film an exploded view, the reassembly and a cutaway in Blender run headless. 1080×1350, 24 fps, 15 seconds. The render must finish inside 30 minutes; our own reference scene took 20.2 minutes on the same laptop.
  • JS: every model gets the same pot file and animates the brew in three.js: heat, pressure in the headspace, water up the funnel and through the grounds, coffee up the column, the boiler emptying as the top fills. The checker films every page with the same software renderer, so a faster GPU path gives no advantage.
  • Same machine for every model: Core Ultra 7 155H laptop, Intel Arc iGPU, 16 GB LPDDR5X, Linux, Blender 5.3, opencode, the model's own provider. One run per model; a rerun is published beside the first, never over it.
  • Real hardware limits, on purpose: no discrete GPU, a 30-minute render budget, and a browser page that has to run smoothly on the integrated graphics. The prompt states all of it, so each model has to plan around it: pick a render engine and sample count, render test frames, measure, then commit. These films are not the best each model could make on a workstation. The test scores spatial reasoning in 3D, and also the judgement a model shows when the hardware says no.

How it is scored (v1.1)

Unforgiving by design: scores are not curved, a bad execution gets a bad number, and every result is published with its reasons.

  • Blender, 100 points: checker 20 (parts in the right place, the right shape, the exploded view and the turn around the pot); judged 70 (construction fidelity 15, camera 15, materials 10, light 10, finish 10, cutaway 10); efficiency 10.
  • JS, 100 points: checker 10 (real-time speed, loop and framing checks); judged 80 (the brew process shown correctly 30, look 30, clarity 15, a clean loop 5); efficiency 10.
  • Following the brief earns nothing. Breaking it costs 5 points per item.
  • A fatal flaw caps the film at 40, for example a pot too dark to read or water that renders as a void. The flaw is printed on the card.
  • Judged items are scored against written anchors at zero, half and full marks, where full marks mean premium-ad quality, and compared across models before a score is final. The judging is not blind: the judge also runs the setup.
  • Efficiency: one prompt scores best, and fewer tokens score more (50K or less is full marks, 300K or more is zero). Tokens stand in for the time a model spends getting there: a 15-minute run is not worth the same as an hour-long one. Efficiency is not cost: providers and plans differ, and price is not scored for now.