Quantization publisher
ibm-granite
ibm-granite publishes 248 quantizations across 22 models in our index, averaging 5.995 effective bits per weight. Their files differ in size from other publishers' builds of the same nominal quantization on 25 of the pairs we can compare — the same label does not mean the same file.
From the file· summed file bytes
Repositories
22
Quantizations
248
Models covered
22
Avg effective bpw
5.995
across their files
Same model, same quant label, different bytes
largest disagreements first
| Model | Quant | ibm-granite | vs | Theirs | Difference |
|---|---|---|---|---|---|
| granite-speech-4.1-2b-plus | BF16 | 3.04 GiB | handy-computer | 3.94 GiB | -22.7% |
| granite-speech-4.1-2b | BF16 | 3.43 GiB | handy-computer | 4.31 GiB | -20.6% |
| granite-4.0-h-small | Q3_K_L | 15.34 GiB | bartowski | 14.50 GiB | +5.8% |
| granite-vision-3.3-2b | Q8_0 | 2.51 GiB | cstr | 3.14 GiB | -20.0% |
| granite-4.0-h-small | Q4_K_S | 17.16 GiB | bartowski | 17.78 GiB | -3.5% |
| granite-speech-4.1-2b-plus | Q8_0 | 1.62 GiB | handy-computer | 2.18 GiB | -25.9% |
| granite-speech-4.1-2b | Q8_0 | 1.82 GiB | handy-computer | 2.38 GiB | -23.6% |
| granite-4.0-h-small | Q4_0 | 17.02 GiB | bartowski | 17.51 GiB | -2.8% |
| granite-speech-4.1-2b-plus | Q6_K | 1.25 GiB | handy-computer | 1.73 GiB | -27.8% |
| granite-speech-4.1-2b | Q6_K | 1.41 GiB | handy-computer | 1.89 GiB | -25.4% |
| granite-speech-4.1-2b-plus | Q5_K_M | 1.10 GiB | handy-computer | 1.58 GiB | -30.3% |
| granite-speech-4.1-2b | Q5_K_M | 1.23 GiB | handy-computer | 1.70 GiB | -27.9% |
| granite-speech-4.1-2b-plus | Q4_K_M | 0.95 GiB | handy-computer | 1.39 GiB | -31.3% |
| granite-speech-4.1-2b | Q4_K_M | 1.06 GiB | handy-computer | 1.49 GiB | -28.9% |
| granite-4.0-h-small | Q3_K_S | 13.09 GiB | bartowski | 13.43 GiB | -2.5% |
| granite-4.0-h-small | Q3_K_M | 14.31 GiB | bartowski | 13.99 GiB | +2.3% |
| granite-4.0-h-small | Q5_K_M | 21.30 GiB | bartowski | 21.55 GiB | -1.2% |
| granite-4.0-h-small | Q4_K_M | 18.14 GiB | bartowski | 18.39 GiB | -1.4% |
| granite-4.0-h-small | Q5_K_S | 20.72 GiB | bartowski | 20.90 GiB | -0.9% |
| granite-4.0-h-small | Q4_1 | 18.87 GiB | bartowski | 19.05 GiB | -1.0% |
| granite-4.0-h-small | Q6_K | 24.65 GiB | bartowski | 24.82 GiB | -0.7% |
| granite-4.0-h-tiny | Q2_K | 2.41 GiB | unsloth | 2.46 GiB | -2.0% |
| granite-4.0-h-tiny | Q4_0 | 3.69 GiB | unsloth | 3.74 GiB | -1.2% |
| granite-4.0-h-small | Q2_K | 10.97 GiB | bartowski | 10.93 GiB | +0.4% |
| granite-4.0-h-tiny | Q3_K_S | 2.85 GiB | unsloth | 2.89 GiB | -1.3% |
A quantization label describes a target, not a recipe. Publishers make different choices about which tensors to keep at higher precision, and some apply an importance matrix while others don't — so two files both honestly labelled the same thing can differ measurably in size and in quality.
Models they publish
| Model | Quantizations | Smallest |
|---|---|---|
| granite-4.1-3b | 15 | 1.28 GiB |
| granite-speech-4.1-2b | 5 | 1.06 GiB |
| granite-4.1-8b | 15 | 3.18 GiB |
| granite-4.0-h-tiny | 15 | 2.41 GiB |
| granite-4.1-30b | 15 | 9.99 GiB |
| granite-speech-4.1-2b-plus | 5 | 0.95 GiB |
| granite-4.0-micro | 15 | 1.28 GiB |
| granite-4.0-h-small | 15 | 10.97 GiB |
| granite-4.0-h-micro | 15 | 1.14 GiB |
| granite-4.0-350m | 15 | 0.17 GiB |
| granite-8b-code-base-4k | 1 | 4.55 GiB |
| granite-3.3-8b-instruct | 15 | 2.89 GiB |
| granite-4.0-1b | 15 | 0.65 GiB |
| granite-34b-code-base-8k | 1 | 19.92 GiB |
| granite-4.0-micro-base | 15 | 1.28 GiB |
| granite-4.0-h-tiny-base | 15 | 2.41 GiB |
| granite-4.0-h-1b | 15 | 0.55 GiB |
| granite-3.3-2b-instruct | 15 | 0.91 GiB |
| granite-4.0-h-350m | 15 | 0.15 GiB |
| granite-vision-4.1-4b | 5 | 1.96 GiB |
| granite-20b-code-base-8k | 1 | 11.94 GiB |
| granite-vision-3.3-2b | 5 | 1.44 GiB |