Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →DistilQwen2.5 — Alibaba's industrial practices for training distilled open lightweight language models. Knowledge distillation from Qwen2.5 72B into smaller 0.5B-7B models.
.claude/skills/mkurman-distilqwen/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -21% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -68% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -52% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -31% | 0% |
DistilQwen2.5 (Alibaba, arXiv:2504.15027) provides industrial practices for training distilled open lightweight LLMs. The approach distills Qwen2.5-72B-Instruct into smaller models (0.5B, 1.5B, 3B, 7B) with strong performance retention.
The distilled models are available on HuggingFace as distilqwen/distilqwen2.5-*-instruct and can be used directly:
pythonfrom transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained("distilqwen/distilqwen2.5-1.5b-instruct") tokenizer = AutoTokenizer.from_pretrained("distilqwen/distilqwen2.5-1.5b-instruct")
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 9,335 | 5,484 | -41% | 1 | 1 | 0% | 1,930 | 1,232 | -36% | 0 | 0 | — |
case-02 | fail→pass | 17,818 | 12,677 | -29% | 1 | 1 | 0% | 3,155 | 2,503 | -21% | 0 | 0 | — |
case-03 | fail→pass | 14,048 | 3,135 | -78% | 1 | 1 | 0% | 2,828 | 913 | -68% | 0 | 0 | — |
case-04 | fail→pass | 11,105 | 3,805 | -66% | 1 | 1 | 0% | 1,923 | 931 | -52% | 0 | 0 | — |
case-05 | pass→pass | 11,934 | 2,062 | -83% | 1 | 1 | 0% | 2,279 | 636 | -72% | 0 | 0 | — |
case-06 | pass→pass | 6,930 | 1,976 | -71% | 1 | 1 | 0% | 1,242 | 613 | -51% | 0 | 0 | — |
case-07 | pass→pass | 10,853 | 2,243 | -79% | 1 | 1 | 0% | 1,717 | 647 | -62% | 0 | 0 | — |
case-08 | pass→pass | 14,662 | 8,253 | -44% | 1 | 1 | 0% | 2,558 | 1,796 | -30% | 0 | 0 | — |
case-09 | fail→pass | 6,003 | 2,479 | -59% | 1 | 1 | 0% | 1,072 | 737 | -31% | 0 | 0 | — |
case-10 | fail→pass | 9,647 | 2,407 | -75% | 1 | 1 | 0% | 1,847 | 760 | -59% | 0 | 0 | — |
case-11 | fail→pass | 8,755 | 2,295 | -74% | 1 | 1 | 0% | 1,616 | 711 | -56% | 0 | 0 | — |
case-12 | pass→pass | 15,492 | 13,141 | -15% | 1 | 1 | 0% | 2,453 | 2,477 | +1% | 0 | 0 | — |
case-13 | pass→pass | 10,560 | 1,993 | -81% | 1 | 1 | 0% | 1,920 | 584 | -70% | 0 | 0 | — |
case-14 | pass→pass | 8,921 | 1,955 | -78% | 1 | 1 | 0% | 1,501 | 573 | -62% | 0 | 0 | — |
case-15 | pass→pass | 11,350 | 2,577 | -77% | 1 | 1 | 0% | 1,791 | 701 | -61% | 0 | 0 | — |
case-16 | pass→pass | 6,966 | 2,205 | -68% | 1 | 1 | 0% | 1,291 | 641 | -50% | 0 | 0 | — |
case-17 | fail→pass | 10,006 | 3,183 | -68% | 1 | 1 | 0% | 1,892 | 969 | -49% | 0 | 0 | — |
case-18 | pass→pass | 8,744 | 5,865 | -33% | 1 | 1 | 0% | 1,559 | 1,335 | -14% | 0 | 0 | — |
case-19 | pass→pass | 11,856 | 3,066 | -74% | 1 | 1 | 0% | 2,251 | 855 | -62% | 0 | 0 | — |
case-20 | pass→pass | 16,545 | 16,171 | -2% | 1 | 1 | 0% | 3,278 | 3,741 | +14% | 0 | 0 | — |
case-21 | pass→pass | 17,568 | 11,025 | -37% | 1 | 1 | 0% | 3,144 | 2,302 | -27% | 0 | 0 | — |
case-22 | pass→pass | 10,874 | 6,726 | -38% | 1 | 1 | 0% | 2,156 | 1,713 | -21% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +36 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.