gemma-4-E2B-it-GMCE (GMCoder Efficiency)
google/gemma-4-E2B-it fine-tuned for GMCoder Efficiency, which makes answers shorter and more direct. Files: GGUF (Q8_0, Q6_K, Q5_K_M, Q4_K_M, BF16) and BF16 safetensors.
What it does (measured)
Our own unseen "Form B" exams, default chat mode, 60k token cap, single run per model, graded blind by one grader. Scores are approximate.
| Tokens (3 exams) | Ultimate (/25) | Historical (/100, approx) | |
|---|---|---|---|
| gemma-4-E2B-it | 20.4k | 17 | ~58 |
| gemma-4-E2B-it-GMCE | 13.5k (-34%) | 10 | ~50 |
Caveats
- Less accurate than the base model in our tests: it is cheaper, but on the hardest exam it stopped after question 20 and scored lower. At this size the recipe costs accuracy. The larger gemma-4-E4B-it-GMCE holds accuracy much better.
- Only the default mode was tested (our local server ignored Gemma's thinking switch).
- Benchmarks are our own and small.