Quantization publisher
mradermacher
mradermacher publishes 31,922 quantizations across 1158 models in our index, averaging 4.345 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
1,802
Quantizations
31,922
Models covered
1158
Avg effective bpw
4.345
across their files
Same model, same quant label, different bytes
largest disagreements first
| Model | Quant | mradermacher | vs | Theirs | Difference |
|---|---|---|---|---|---|
| Aurora-Code-1 | Q6_K | 23.37 GiB | bartowski | 27.99 GiB | -16.5% |
| Aurora-Code-1 | Q8_0 | 30.25 GiB | bartowski | 34.38 GiB | -12.0% |
| deepseek-llm-67b-chat | Q2_K | 23.40 GiB | TheBloke | 26.54 GiB | -11.8% |
| Aurora-Code-1 | Q5_K_M | 20.23 GiB | bartowski | 23.30 GiB | -13.2% |
| gemma-4-E4B-it-The-DECKARD-Expresso-Universe-HERETIC-UNCENSORED-Thinking | Q8_0 | 7.48 GiB | DavidAU | 10.53 GiB | -28.9% |
| Gemma-4-26B-A4B-StyleTune-V2 | Q4_K_M | 16.03 GiB | EmanuelOverride | 13.04 GiB | +22.9% |
| Aurora-Code-1 | Q4_K_S | 16.26 GiB | bartowski | 19.18 GiB | -15.2% |
| Aurora-Code-1 | Q5_K_S | 19.63 GiB | bartowski | 22.50 GiB | -12.7% |
| Aurora-Code-1 | Q4_K_M | 17.28 GiB | bartowski | 19.92 GiB | -13.3% |
| mythos-9b-unhinged | Q5_K_S | 5.33 GiB | fableforge-ai | 3.20 GiB | +66.8% |
| Aurora-Code-1 | IQ4_XS | 15.42 GiB | bartowski | 17.51 GiB | -12.0% |
| Aurora-Code-1 | Q3_K_S | 12.38 GiB | bartowski | 14.45 GiB | -14.3% |
| Mixtral-8x7B-Instruct-v0.1 | Q3_K_M | 21.00 GiB | TheBloke | 18.96 GiB | +10.7% |
| dolphin-2.6-mixtral-8x7b | Q3_K_M | 21.00 GiB | TheBloke | 18.96 GiB | +10.7% |
| Mixtral-8x7B-Instruct-v0.1 | Q4_K_M | 26.49 GiB | TheBloke | 24.63 GiB | +7.6% |
| dolphin-2.6-mixtral-8x7b | Q4_K_M | 26.49 GiB | TheBloke | 24.63 GiB | +7.6% |
| ReasonCritic-7B | F16 | 15.26 GiB | fableforge-ai | 13.44 GiB | +13.6% |
| Capybara-Tess-Yi-34B-200K | Q2_K | 11.94 GiB | TheBloke | 13.56 GiB | -11.9% |
| WizardCoder-Python-34B-V1.0 | Q2_K | 11.65 GiB | TheBloke | 13.23 GiB | -12.0% |
| deepseek-coder-33b-instruct | Q2_K | 11.51 GiB | TheBloke | 13.07 GiB | -11.9% |
| G4-MeroMero-26B-A4B-it-uncensored-heretic | Q3_K_M | 12.37 GiB | llmfan46 | 13.93 GiB | -11.2% |
| Mixtral-8x7B-Instruct-v0.1 | Q2_K | 16.12 GiB | TheBloke | 14.57 GiB | +10.7% |
| dolphin-2.6-mixtral-8x7b | Q2_K | 16.12 GiB | TheBloke | 14.57 GiB | +10.7% |
| G4-MeroMero-26B-A4B-it-uncensored-heretic | Q3_K_L | 12.88 GiB | llmfan46 | 14.37 GiB | -10.4% |
| grug-35b-v2 | Q6_K | 26.56 GiB | bartowski | 27.99 GiB | -5.1% |
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.