Underdog Saluki 27B IQ2-mix (llama.cpp)

Metrics

Checklist score

– / 100

Design (rubric, 0–5)

–

Stop reason

done

Model time

31mins

Tool time (wall minus model)

13s

Wall time

31mins

Tool calls

23

Tool errors

4

Errors by type

{"other": 1, "format": 3}

Loop-guard events

0

Tokens in / out

1,140,958.0 / 25,863.0

Thinking tokens

–

Decode / prefill tok/s

15.1 / 378.5

Prefix reuse

0.960

Peak VRAM (GPU / engine)

10,367 / 10,358 MiB

Peak engine RAM

10,944 MiB

Engine

llamacpp

Model

Underdog Saluki 27B IQ2-mix (llama.cpp) (saluki-27b)

Settings

{"thinking": "on (Qwen3.8 template default), server --reasoning-budget 3072", "reasoning_budget_tokens": 3072, "max_tokens": 7168, "sampling": "Saluki card thinking recipe: temp 0.6, top_p 0.95, top_k 20, min_p 0", "context": "65536 (q8_0 KV), 32768 fallback"}

Node

omarchy (RTX 3060 12 GB)

Prompt

v3

Started

2026-10-11T01:34:44Z

Feature scores