Nanbeige4.2-3B-GMCE (GMCoder Efficiency)
Nanbeige4.2-3B fine-tuned for GMCoder Efficiency, which pushes the model toward shorter, more direct answers. File: Nanbeige4.2-3B-GMCE.Q8_0.gguf (Q8_0).
Usage: keep thinking ON (the default)
This model is built for the model's normal thinking mode, where it thinks noticeably less than the base. Do not switch thinking off: in our tests answer quality drops sharply with thinking off (Form B Historical about 69.5 with thinking on vs about 46 with thinking off).
What it does (measured)
Single run per model, one grader (blind to model identity), so a few points either way is noise. Default thinking mode, 40k-60k token cap, our own unseen "Form B" exams (275 points total).
| Tokens (3 exams) | Score | |
|---|---|---|
| Nanbeige4.2-3B base | 57.7k | ~208 / 275 |
| Nanbeige4.2-3B-GMCE | 46.8k (-19%) | ~195 / 275 |
Per exam, token savings were uneven: Historical -54%, Everyday -29%, Ultimate (math/logic-heavy) about 0%. On a separate quick 10-task coding check (8k cap) it used 18% fewer tokens with the same pass rate.
Caveats
- This is a modest efficiency gain with a small accuracy cost (about 13 points on Form B). Most of Nanbeige's tokens are thinking, and this recipe only trims it.
- GMCoder 9B with the same recipe saved about 52% of tokens at a smaller accuracy cost (-12 of 275).
- Benchmarks are our own and small; do not read the numbers as leaderboard results.