Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Tencent AngelSlim — accessible, comprehensive, and efficient toolkit for large model compression. Quantization (FP8/INT4/NVFP4/1.25-bit), pruning, speculative decoding (Eagle3), and diffusion model compression.
.claude/skills/mkurman-angelslim/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -17% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -17% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -64% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -71% | 0% |
AngelSlim integrates mainstream compression algorithms into a unified framework with one-click access. Supports FP8/INT8/INT4/NVFP4/1.25-bit quantization, pruning, Eagle3 speculative decoding, and diffusion model compression for LLMs, VLMs, and audio models.
bashuv pip install angelslim
bashimport angelslim as slim # FP8 static quantization model = slim.quantize(model, dtype="fp8_static", qconfig="default") # INT4 GPTQ model = slim.quantize(model, dtype="int4_gptq", dataset="wikitext2")
| Method | Precision | Best For | |---|---|---| | FP8-Static/Dynamic | 8-bit | General LLM deployment | | INT4 GPTQ/AWQ/GPTAQ | 4-bit | Memory-constrained serving | | NVFP4 | 4-bit (NVIDIA) | Blackwell GPUs | | Sherry | 1.25-bit | Extreme compression | | STQ1_0 | 1.25-bit | On-device deployment |
python# Train Eagle3 draft model slim.eagle3.train(model, draft_model_config) # Inference with Eagle3 output = model.generate_with_eagle3(input_ids, max_new_tokens=256)
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-05 | pass→pass | 9,969 | 2,757 | -72% | 1 | 1 | 0% | 1,644 | 764 | -54% | 0 | 0 | — |
case-01 | fail→pass | 10,019 | 5,786 | -42% | 1 | 1 | 0% | 1,925 | 1,600 | -17% | 0 | 0 | — |
case-02 | fail→fail | 15,671 | 9,608 | -39% | 1 | 1 | 0% | 2,827 | 2,295 | -19% | 0 | 0 | — |
case-03 | fail→pass | 8,792 | 3,593 | -59% | 1 | 1 | 0% | 1,198 | 992 | -17% | 0 | 0 | — |
case-04 | fail→pass | 9,465 | 1,632 | -83% | 1 | 1 | 0% | 1,312 | 593 | -55% | 0 | 0 | — |
case-06 | fail→pass | 18,345 | 3,731 | -80% | 1 | 1 | 0% | 2,815 | 1,023 | -64% | 0 | 0 | — |
case-07 | fail→pass | 23,086 | 3,507 | -85% | 1 | 1 | 0% | 3,517 | 1,025 | -71% | 0 | 0 | — |
case-08 | fail→pass | 12,854 | 3,020 | -77% | 1 | 1 | 0% | 2,106 | 961 | -54% | 0 | 0 | — |
case-09 | pass→pass | 33,999 | 4,656 | -86% | 1 | 1 | 0% | 2,884 | 1,265 | -56% | 0 | 0 | — |
case-10 | fail→pass | 14,534 | 3,557 | -76% | 1 | 1 | 0% | 2,298 | 978 | -57% | 0 | 0 | — |
case-11 | fail→pass | 10,051 | 3,350 | -67% | 1 | 1 | 0% | 1,692 | 581 | -66% | 0 | 0 | — |
case-12 | pass→pass | 14,378 | 2,668 | -81% | 1 | 1 | 0% | 2,127 | 829 | -61% | 0 | 0 | — |
case-13 | pass→pass | 11,389 | 3,796 | -67% | 1 | 1 | 0% | 1,873 | 857 | -54% | 0 | 0 | — |
case-14 | pass→pass | 13,694 | 2,703 | -80% | 1 | 1 | 0% | 1,910 | 787 | -59% | 0 | 0 | — |
case-15 | fail→pass | 11,047 | 1,711 | -85% | 1 | 1 | 0% | 1,681 | 633 | -62% | 0 | 0 | — |
case-16 | pass→pass | 4,119 | 1,875 | -54% | 1 | 1 | 0% | 530 | 639 | +21% | 0 | 0 | — |
case-17 | fail→pass | 6,494 | 2,133 | -67% | 1 | 1 | 0% | 1,051 | 793 | -25% | 0 | 0 | — |
case-18 | fail→pass | 14,620 | 1,792 | -88% | 1 | 1 | 0% | 2,443 | 640 | -74% | 0 | 0 | — |
case-19 | fail→pass | 5,462 | 1,824 | -67% | 1 | 1 | 0% | 783 | 700 | -11% | 0 | 0 | — |
case-20 | pass→pass | 14,856 | 8,468 | -43% | 1 | 1 | 0% | 2,738 | 1,962 | -28% | 0 | 0 | — |
case-21 | pass→pass | 23,193 | 19,210 | -17% | 1 | 1 | 0% | 4,400 | 4,112 | -7% | 0 | 0 | — |
case-22 | pass→pass | 13,565 | 8,945 | -34% | 1 | 1 | 0% | 2,335 | 2,070 | -11% | 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 +55 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.