Quantization format
BF16
Files labelled BF16 average 15.545 effective bits per weight across 945 real quantizations — not the nominal 16. That is -3% more than the label implies, because a quantization is a mixture: some tensors are always kept at higher precision.
From the file· 945 files measured
Nominal bpw
16
from the block layout
Measured average
15.545
945 files
Range
1.033–32.883
varies by architecture
File sizes
0.05 GiB+
up to 1912.15 GiB
Note
Unquantized bfloat16; wider exponent, fewer mantissa bits than F16.
Real files
one per model, most downloaded first
| Model | Size● | Effective bpw● | vs nominal● |
|---|---|---|---|
| Qwen3.6-27B | 50.11 GiB | 15.495 | +-3% |
| embeddinggemma-300m | 0.57 GiB | 16.177 | +1% |
| Qwen3.6-35B-A3B | 64.61 GiB | 15.438 | +-4% |
| Qwen3.5-9B | 16.69 GiB | 14.852 | +-7% |
| gemma-4-26B-A4B-it | 47.04 GiB | 15.222 | +-5% |
| gemma-4-12B-it | 22.20 GiB | 15.941 | +-0% |
| Qwen3.5-4B | 7.85 GiB | 14.463 | +-10% |
| Qwythos-9B-Claude-Mythos-5-1M | 17.14 GiB | 15.649 | +-2% |
| gemma-4-E4B-it | 14.02 GiB | 15.060 | +-6% |
| Qwen3-Coder-30B-A3B-Instruct | 56.90 GiB | 16.008 | +0% |
| gemma-4-31B-it | 57.20 GiB | 15.710 | +-2% |
| Qwen3-VL-30B-A3B-Instruct | 56.90 GiB | 15.731 | +-2% |
| FLUX.2-klein-9B | 16.91 GiB | 16.000 | +0% |
| cohere-transcribe-03-2026 | 3.82 GiB | 15.898 | +-1% |
| Qwen-AgentWorld-35B-A3B | 64.61 GiB | 16.013 | +0% |
| LTX-2.3 | 39.15 GiB | — | — |
| Llama-3.2-1B-Instruct | 2.31 GiB | 16.052 | +0% |
| gemma-4-E2B-it | 8.67 GiB | 14.540 | +-9% |
| Qwen3-8B | 15.26 GiB | 16.006 | +0% |
| Qwen3.5-0.8B | 1.41 GiB | 13.892 | +-13% |
| Qwen3-4B | 7.50 GiB | 16.013 | +0% |
| GLM-5.2 | 1404.42 GiB | 16.014 | +0% |
| Llama-3.1-8B-Instruct | 14.97 GiB | 16.008 | +0% |
| Ornith-1.0-35B | 64.61 GiB | 16.013 | +0% |
| Llama-3.2-3B-Instruct | 5.99 GiB | 16.020 | +0% |
●From the filewhat these mean