K2-Horizon-3.7B-GMCE (GMCoder Efficiency)
IFM/K2-Horizon-3.7B fine-tuned for GMCoder Efficiency, which cuts how many tokens the model spends reasoning before it answers. 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 much 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, 60k token cap. Compared against the base at its default setting and at medium reasoning effort.
| Historical tokens | Everyday tokens | Historical score (/100, approx) | |
|---|---|---|---|
| K2-Horizon-3.7B base, default | 14.1k | 12.0k | ~78 |
| K2-Horizon-3.7B base, medium effort | 5.1k | 6.3k | ~68 |
| K2-Horizon-3.7B-GMCE | 5.9k (-58%) | 4.6k (-62%) | ~75 |
On the hardest exam (Ultimate, 25 math/logic questions) the base at default, the base at medium effort and this model all ran to the 60k cap without answering.
Caveats
- A large token saving against the default setting for a few points of accuracy.
- The comparison with medium effort is within noise on a single run.
- Benchmarks are our own and small. Do not read the numbers as leaderboard results.