Skip to content

All model benchmarks

Holds back, game tasks

Tryll game suites, one greedy pass

Only the game-task lines where no tool fits, such as small talk next to a door the player could open. A line counts when the model makes no call. A model that never calls anything scores full marks here, so read it beside Tool calls, game tasks.

Every model on this page was measured on NVIDIA GeForce RTX 4080 · Vulkan.

Holds back, game tasks

Holds back, game tasks Holds back, game tasks

Measured models

Model Right Interval Rank range Cases Family
Bonsai 8B (Q1_0 1-bit) 100.0% 100.0% to 100.0% #1–10 of 12 24 of 24 Bonsai
Bonsai 8B (Q2_0 2-bit) 100.0% 100.0% to 100.0% #1–10 of 12 24 of 24 Bonsai
Gemma 4 E2B Instruct QAT (UD-Q4_K_XL) 100.0% 100.0% to 100.0% #1–10 of 12 24 of 24 Gemma
Gemma 4 E4B Instruct (Q4_K_M) 100.0% 100.0% to 100.0% #1–10 of 12 24 of 24 Gemma
Granite 4.2 3B Instruct (Q4_K_M) 100.0% 100.0% to 100.0% #1–10 of 12 24 of 24 Granite
Granite 4.2 8B Instruct (Q4_K_M) 100.0% 100.0% to 100.0% #1–10 of 12 24 of 24 Granite
Ministral 3 8B Instruct (Q4_K_M) 100.0% 100.0% to 100.0% #1–10 of 12 24 of 24 Ministral
Ministral 3 3B Instruct (Q4_K_M) 95.8% 87.5% to 100.0% #1–10 of 12 23 of 24 Ministral
Ling 3.0 Tiny (Q4_K_M) 91.7% 79.2% to 100.0% #1–10 of 12 22 of 24 Ling
Qwen 3.5 4B (Q4_K_M) 87.5% 75.0% to 100.0% #1–10 of 12 21 of 24 Qwen
Llama 3.1 8B Instruct (Q4_K_M) 25.0% 8.3% to 41.7% #11–12 of 12 6 of 24 Llama
Llama 3.2 3B Instruct (Q4_K_M) 8.3% 0.0% to 20.8% #11–12 of 12 2 of 24 Llama

Model cards

No tool support: SmolLM2 1.7B Instruct (Q4_K_M).