Quantization format
F32
Files labelled F32 average 31.257 effective bits per weight across 158 real quantizations — not the nominal 32. That is -2% more than the label implies, because a quantization is a mixture: some tensors are always kept at higher precision.
From the file· 158 files measured
Nominal bpw
32
from the block layout
Measured average
31.257
158 files
Range
5.061–32.414
varies by architecture
File sizes
0.08 GiB+
up to 262.84 GiB
Note
Full precision. Rarely distributed for inference.
Real files
one per model, most downloaded first
| Model | Size● | Effective bpw● | vs nominal● |
|---|---|---|---|
| embeddinggemma-300m | 1.13 GiB | 32.172 | +1% |
| nemotron-3.5-asr-streaming-0.6b | 2.38 GiB | 32.004 | +0% |
| parakeet-unified-en-0.6b | 2.30 GiB | 32.001 | +0% |
| Llama-3.1-8B-Instruct | 29.92 GiB | 32.008 | +0% |
| parakeet-tdt-0.6b-v3 | 2.34 GiB | 32.003 | +0% |
| whisper-medium | 2.85 GiB | 32.021 | +0% |
| canary-180m-flash | 0.70 GiB | 32.007 | +0% |
| Qwen2.5-3B-Instruct | 11.50 GiB | 32.015 | +0% |
| Phi-3.5-mini-instruct | 14.24 GiB | 32.002 | +0% |
| gemma-2-2b-it | 9.74 GiB | 32.019 | +0% |
| Mistral-Nemo-Instruct-2407 | 45.63 GiB | 32.005 | +0% |
| parakeet-tdt-0.6b-v2 | 2.30 GiB | 32.001 | +0% |
| DeepSeek-R1-Distill-Qwen-7B | 28.38 GiB | 32.006 | +0% |
| nomic-embed-text-v1.5 | 0.51 GiB | 32.043 | +0% |
| DeepSeek-R1-Distill-Qwen-14B | 55.03 GiB | 32.003 | +0% |
| Mistral-Small-24B-Instruct-2501 | 87.82 GiB | 32.003 | +0% |
| GigaAM-v3 | 0.82 GiB | 31.767 | +-1% |
| Mistral-7B-Instruct-v0.3 | 27.00 GiB | 32.001 | +0% |
| umt5-xxl | 21.17 GiB | 32.009 | +0% |
| VibeThinker-3B | 11.50 GiB | 32.015 | +0% |
| Qwen3-TTS-12Hz-0.6B-Base | 28.88 GiB | — | — |
| phi-4 | 54.61 GiB | 32.002 | +0% |
| canary-1b-v2 | 3.65 GiB | 32.004 | +0% |
| DeepSeek-R1-Distill-Qwen-1.5B | 6.63 GiB | 32.027 | +0% |
| Qwen2-7B-Instruct | 28.38 GiB | 32.006 | +0% |
●From the filewhat these mean