Qwen 3.6 27B NVFP4 Thinking | Benchmark #214

Described in:Icy Tower-like game built by Qwen 3.6 and Qwen 3.8 as agents

Benchmark summary

Icy Tower-like game built in SC# as an agent

ModelQwen 3.6 27B QuantizationNVFP4 ModeThinking

The model builds an Icy Tower-like tower jumper game in SC# as an agent: it writes the app, compiles it, runs its own tests and fixes the code. Hidden tests check the rules from the task; each hidden test is its own result.

Score 14/14 100.0% passed
14 passed14 total 1 AI query 5 tool calls TOKENS IN141 723 TOKENS OUT50 515

Results by difficulty

1 group
Level Unknown / other
14 / 14 100.0%
passed
1 AI query 5 tool calls IN141 723 OUT50 515

Results by category

1 group
Category Other
100.0% 14/14 passed
1 AI query 5 tool calls TOKENS IN141 723 TOKENS OUT50 515

Tests with multiple category tags are counted in each matching category, so category totals may exceed the report total.

Benchmark items
Script / test Status Tokens
Building Games in SC#: Icy Tower-like tower jumper 14 passed
100.0%
IN141 723 OUT50 515
✓
IcyTowerThe model builds an Icy Tower style game from the task as an agent; this item passes when the app compiles and can be exported, and each hidden test is its own item.1 AI query 5 tool calls
Passed IN141 723 OUT50 515
✓
Hidden_CameraFollowsThe camera keeps the top of the player at least 60 px below the top of the screen.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_ComboTwo or more combo jumps of at least 2 floors each add (floors of the combo)^2 when the combo ends with a short jump.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_ComboTimerThe combo ends when its timer runs out, 90 frames after the last combo jump.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_FloorsGrowThe floor lists grow with the player and keep following the floor rule.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_GameOverAndRestartA player whose top goes below the bottom of the screen loses; nothing moves then, and Enter starts a new game.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_InitialStateAfter AppInit the player stands on the ground, the counters have their starting values, and the floors follow the floor rule.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_JumpingA jump starts with vy = -(8 + 0.5 * |vx|), only from a floor; gravity adds 0.5 per frame up to 10.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_LandingA falling player lands on a floor it overlaps; it jumps through floors from below.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_RunningLeft and Right speed the player up by 0.4 per frame up to 6; without keys the speed shrinks by 15 % per frame and stops below 0.05.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_ScreenShowsTheTowerThe screen shows the walls, the floors on the screen in their colors, the player and the texts. Texts and part B drawings may cover a few pixels, so the pixel counts allow a small loss but nothing extra.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_ScrollingScrolling starts at floor 5, moves the camera up by scrollSpeed every frame and gets 0.5 faster every 900 frames, up to 5.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_WalkingOffA player who walks off the end of a floor falls, and lands on a lower floor.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0
✓
Hidden_WallBounceThe player bounces off a wall with three quarters of the speed in the other direction.0 AI queries 0 tool callsHiddenTest
Passed IN0 OUT0