Quantization format
Q3_K_S
Files labelled Q3_K_S average 3.744 effective bits per weight across 2241 real quantizations — not the nominal 3.4375. That is 9% more than the label implies, because a quantization is a mixture: some tensors are always kept at higher precision.
From the file· 2241 files measured
Nominal bpw
3.4375
from the block layout
Measured average
3.744
2241 files
Range
1.403–21.256
varies by architecture
File sizes
0.02 GiB+
up to 414.28 GiB
Real files
one per model, most downloaded first
| Model | Size● | Effective bpw● | vs nominal● |
|---|---|---|---|
| Qwen3.6-27B | 11.51 GiB | 3.559 | +4% |
| Qwen3.6-35B-A3B | 15.28 GiB | 3.650 | +6% |
| Qwen3.5-9B | 4.02 GiB | 3.578 | +4% |
| gemma-4-26B-A4B-it | 11.69 GiB | 3.783 | +10% |
| gemma-4-12B-it | 4.78 GiB | 3.435 | +-0% |
| Qwen3.5-4B | 1.96 GiB | 3.615 | +5% |
| Hy3 | 122.26 GiB | 3.515 | +2% |
| gemma-4-E4B-it | 3.60 GiB | 3.864 | +12% |
| Qwen3-Coder-30B-A3B-Instruct | 12.38 GiB | 3.483 | +1% |
| gemma-4-31B-it | 12.30 GiB | 3.380 | +-2% |
| Qwen3-VL-30B-A3B-Instruct | 12.38 GiB | 3.422 | +-0% |
| FLUX.2-klein-9B | 4.37 GiB | 4.132 | +20% |
| LTX-2.3 | 9.07 GiB | 8.474 | +147% |
| Llama-3.2-1B-Instruct | 0.60 GiB | 4.154 | +21% |
| gemma-4-E2B-it | 2.28 GiB | 3.819 | +11% |
| gpt-oss-20b | 10.68 GiB | 4.263 | +24% |
| Qwen3-8B | 3.51 GiB | 3.682 | +7% |
| Qwen3.5-0.8B | 0.41 GiB | 4.037 | +17% |
| Qwen3-VL-8B-Instruct-abliterated-v1 | 3.51 GiB | 3.440 | +0% |
| UI-TARS-1.5-7B | 3.25 GiB | 3.369 | +-2% |
| Qwen3-4B | 1.76 GiB | 3.753 | +9% |
| Llama-3.1-8B-Instruct | 3.41 GiB | 3.651 | +6% |
| Ornith-1.0-35B | 14.45 GiB | 3.581 | +4% |
| Llama-3.2-3B-Instruct | 1.44 GiB | 3.842 | +12% |
| ThinkingCap-Qwen3.6-27B | 12.78 GiB | 4.012 | +17% |
●From the filewhat these mean