Qwen3.5-4B-GMCE (GMCoder Efficiency)
Qwen3.5-4B fine-tuned for GMCoder Efficiency, which teaches the model to answer directly and concisely instead of thinking for thousands of tokens. File: Qwen3.5-4B-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 far less than the base. Do not switch thinking off: in our tests answer quality drops sharply with thinking off (Form B Historical about 51 with thinking on vs about 43 with thinking off, and it starts inventing details).
What it does (measured)
Everything below is a single run per model, graded by one grader (blind to model identity), so differences of a few points are noise. Thinking mode was left at the model's default, 40k to 60k token cap.
Unseen "Form B" coding/reasoning exams (3 exams, 275 points total):
| Tokens (3 exams) | Result | |
|---|---|---|
| Qwen3.5-4B base | 109k+ | Did not produce an answer on 2 of 3 exams (ran to the token cap while still thinking) |
| Qwen3.5-4B-GMCE | 21.6k | Finished all three, about 141 / 275 |
Quick coding check (10 hidden-test Python tasks + 3 explain prompts, 8k cap): average tokens per answer 2,915 → 1,479 (-49%), answers hitting the cap 19% → 0%, tasks passed 45% → 45%.
Caveats
- It is much cheaper to run, but it is not a strong coder in absolute terms: about 51% on Form B, versus 76% for GMCoder 9B.
- The recipe reduced length a lot on this model; on smaller Qwen3.5 models (0.8B, 2B) and on Nanbeige it helped little.
- Benchmarks are our own and small. Do not read the numbers as leaderboard results.