Rank #13
Model result · rank #13
Gemma4‑12B‑Coder Fable5/Composer2.5 Q4_K_M
Local modelLocal GGUF · llama.cpp Vulkan · Q4_K_M. Public result card with the model’s overall score, lane measurements, runtime/cost telemetry, and ranking formula.
Full rank #11
SWE rank #13
Hard rank #11
100.0% reliability
All-around publication view
Score61.19
Formulamean(Full, SWE, Hard Intelligence)
BasisLocal GGUF · llama.cpp Vulkan · Q4_K_M
The overall score averages the measured major lanes while keeping each source measurement visible.
Full / Agentic benchmark
Final80.60
Capability87.07
Agentic79.75
Pass rate88.4%
Prompts43
This lane captures instruction following, structured behavior, tool discipline, and general agentic reliability.
Software engineering MVP
SWE score54.01
Focused final54.01
Capability32.70
Daily driver60.76
Prompts8
This lane is closer to implementation usefulness: source handling, architecture cleanliness, and deliverable quality.
Hard Intelligence diagnostic
Hard score48.96
Active inquiry17.50
Online adaptation6.25
Self-repair82.08
Authority integrity90.00
Hard Intelligence measures active inquiry, online adaptation, evidence-driven self-repair, and authority/salience integrity.
Runtime economics
Total cost$0
Cost / scored item$0
Seconds / timed item21.52s
Runtime coverage100.0%
Recorded tokens / item8.3k
Token coverage100.0%
Cost, time, and token basis are normalized telemetry. They explain tradeoffs; they do not overwrite the capability score yet.
Interpretation
Why this result lands here.
The model is stronger in the Full/Agentic lane than in the SWE lane; the overall score is therefore shown with both component lanes visible. Hard Intelligence score is 48.96 and contributes to the overall score alongside Full/Agentic and SWE. Local model row: benchmarked on local hardware with no API metering. Strong local Full/Agentic and focused tool-loop behavior; SWE review/audit and Hard Intelligence inquiry/adaptation are weak in this max-thinking run.