Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Intel Neural Compressor — SOTA low-bit LLM quantization (INT8/FP8/INT4/NVFP4), sparsity, pruning, and distillation for PyTorch, TensorFlow, and ONNX Runtime.
.claude/skills/mkurman-intel-neural-compressor/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -42% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -66% | 0% |
| case-19 | ✗→✓ | ▲ Improved | -9% | 0% |
Intel Neural Compressor provides low-bit quantization (INT8, FP8, INT4, MXFP4, NVFP4), sparsity, pruning, and knowledge distillation for optimizing models on Intel hardware and beyond.
bashuv pip install neural-compressor
pythonfrom neural_compressor import Quantization, config # Post-training quantization quantizer = Quantization(config) q_model = quantizer(model) q_model.save("quantized_model")
pythonfrom neural_compressor import Pruning pruner = Pruning(model, config={"pruning_type": "snip_momentum", "target_sparsity": 0.3}) pruned_model = pruner.fit()
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 9,305 | 6,940 | -25% | 1 | 1 | 0% | 1,844 | 1,580 | -14% | 0 | 0 | — |
case-02 | fail→fail | 13,084 | 9,681 | -26% | 1 | 1 | 0% | 2,417 | 2,147 | -11% | 0 | 0 | — |
case-03 | pass→pass | 12,686 | 7,554 | -40% | 1 | 1 | 0% | 2,428 | 1,658 | -32% | 0 | 0 | — |
case-04 | fail→fail | 7,592 | 6,346 | -16% | 1 | 1 | 0% | 1,454 | 1,340 | -8% | 0 | 0 | — |
case-05 | fail→pass | 11,413 | 7,753 | -32% | 1 | 1 | 0% | 2,132 | 1,727 | -19% | 0 | 0 | — |
case-06 | pass→pass | 4,303 | 1,761 | -59% | 1 | 1 | 0% | 660 | 488 | -26% | 0 | 0 | — |
case-07 | fail→pass | 10,368 | 5,265 | -49% | 1 | 1 | 0% | 1,770 | 1,118 | -37% | 0 | 0 | — |
case-08 | fail→pass | 5,921 | 2,341 | -60% | 1 | 1 | 0% | 1,002 | 583 | -42% | 0 | 0 | — |
case-09 | pass→pass | 5,646 | 2,072 | -63% | 1 | 1 | 0% | 984 | 546 | -45% | 0 | 0 | — |
case-10 | pass→fail | 8,246 | 1,882 | -77% | 1 | 1 | 0% | 1,462 | 486 | -67% | 0 | 0 | — |
case-11 | pass→pass | 9,593 | 2,027 | -79% | 1 | 1 | 0% | 1,606 | 540 | -66% | 0 | 0 | — |
case-12 | fail→pass | 7,955 | 1,856 | -77% | 1 | 1 | 0% | 1,454 | 497 | -66% | 0 | 0 | — |
case-13 | pass→pass | 3,477 | 1,311 | -62% | 1 | 1 | 0% | 532 | 374 | -30% | 0 | 0 | — |
case-14 | pass→pass | 5,072 | 3,151 | -38% | 1 | 1 | 0% | 1,013 | 845 | -17% | 0 | 0 | — |
case-15 | pass→pass | 6,813 | 2,941 | -57% | 1 | 1 | 0% | 1,208 | 696 | -42% | 0 | 0 | — |
case-16 | pass→pass | 5,226 | 1,645 | -69% | 1 | 1 | 0% | 875 | 432 | -51% | 0 | 0 | — |
case-17 | pass→pass | 10,234 | 2,054 | -80% | 1 | 1 | 0% | 1,585 | 465 | -71% | 0 | 0 | — |
case-18 | pass→pass | 8,648 | 2,667 | -69% | 1 | 1 | 0% | 1,703 | 706 | -59% | 0 | 0 | — |
case-19 | fail→pass | 9,164 | 6,786 | -26% | 1 | 1 | 0% | 1,731 | 1,581 | -9% | 0 | 0 | — |
case-20 | pass→pass | 13,747 | 12,413 | -10% | 1 | 1 | 0% | 2,823 | 2,759 | -2% | 0 | 0 | — |
case-21 | pass→pass | 10,660 | 6,715 | -37% | 1 | 1 | 0% | 2,303 | 1,506 | -35% | 0 | 0 | — |
case-22 | pass→pass | 10,411 | 7,175 | -31% | 1 | 1 | 0% | 2,176 | 1,621 | -26% | 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 +18 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
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.