K2-Horizon-7B-GMCE (GMCoder Efficiency)
IFM/K2-Horizon-7B fine-tuned for GMCoder Efficiency, which cuts how many tokens the model spends before answering. Files: GGUF (Q8_0, Q6_K, Q5_K_M, Q4_K_M, BF16) and BF16 safetensors.
GGUF needs llama.cpp release b11454 or newer (K2 Horizon support). K2 Horizon always reasons before answering; this model still does, just far less.
What it does (measured)
Single run per model, one grader (blind to model identity), approximate scores, so differences of a few points are noise. Our own unseen "Form B" exams, default thinking, 60k token cap.
| Exam | Base tokens | GMCE tokens | Saved | Score base → GMCE (approx) |
|---|---|---|---|---|
| Historical (20 coding-reasoning questions, /100) | 14.7k | 4.8k | -67% | ~88 → ~80 |
| Everyday (30 coding questions, /150) | 14.1k | 4.9k | -65% | ~128 → ~112 |
| Ultimate (25 hard math/logic questions) | 60k (cap, no answer) | 60k (cap, no answer) | n/a | neither finished |
Caveats
- A large cut in tokens, with a modest accuracy cost (roughly 8 to 16 points on the two exams it finished).
- On the hardest exam (Ultimate) both the base and this model ran to the 60k cap without answering, so efficiency there is unchanged.
- Benchmarks are our own and small. Do not read the numbers as leaderboard results.