Independent tests of open local models on home-level hardware. Not a taOS feature demo.

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

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

hermes-salukillamacppomarchy (RTX 3060 12 GB, dual-GPU host)scene

This run produced no playable build.

Metrics

Checklist score
– / 100
Design (rubric, 0–5)
–
Stop reason
stall
Model time
33mins
Tool time (wall minus model)
11s
Wall time
33mins
Tool calls
1
Tool errors
1
Errors by type
{"other": 1}
Loop-guard events
0
Tokens in / out
31,493.0 / 34,769.0
Thinking tokens
–
Decode / prefill tok/s
17.6 / 457.8
Prefix reuse
0.747
Peak VRAM (GPU / engine)
3 / 10,354 MiB
Peak engine RAM
9,196 MiB
Engine
llamacpp
Model
Underdog Saluki 27B IQ2-mix (llama.cpp) (saluki-27b)
Settings
{"thinking": "medium", "reasoning_budget_tokens": 3072, "max_tokens": 7168, "sampling": "card: temp 0.6 top_p 0.95 top_k 20 (server)"}
Node
omarchy (RTX 3060 12 GB, dual-GPU host)
Prompt
scene-pagoda-v1
Started
2026-10-30T02:38:16Z

Feature scores

Screenshots

Landing page at desktop and phone widths, and the game at noon, at night, after breaking a block, and on a phone.

Checklist (0 pass, 0 partial, 0 fail)

FeatureTestScoreEvidence

Suite · · WebGL:

Steers

None: an unsteered run gets the prompt and nothing else.