Quantization publisher

backyardai

backyardai publishes 99 quantizations across 5 models in our index, averaging 4.522 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
5
Quantizations
99
Models covered
5
Avg effective bpw
4.522
across their files

Same model, same quant label, different bytes

largest disagreements first
ModelQuantbackyardaivsTheirsDifference
Midnight-Miqu-70B-v1.5IQ4_XS34.30 GiBmradermacher34.64 GiB-1.0%
MN-Violet-Lotus-12BIQ4_XS6.28 GiBmradermacher6.33 GiB-0.8%
Hermes-3-Llama-3.1-70BIQ2_XXS17.79 GiBlegraphista17.79 GiB0.0%
Hermes-3-Llama-3.1-70BIQ4_XS35.30 GiBbartowski35.30 GiB0.0%
Hermes-3-Llama-3.1-70BIQ2_M22.46 GiBlegraphista22.46 GiB0.0%
Hermes-3-Llama-3.1-70BIQ2_XXS17.79 GiBbartowski17.79 GiB0.0%
Hermes-3-Llama-3.1-70BIQ3_XXS25.58 GiBbartowski25.58 GiB0.0%
Hermes-3-Llama-3.1-70BIQ2_XS19.69 GiBbartowski19.69 GiB0.0%
Hermes-3-Llama-3.1-70BIQ2_XS19.69 GiBlegraphista19.69 GiB0.0%
Hermes-3-Llama-3.1-70BIQ4_XS35.30 GiBlegraphista35.30 GiB0.0%
Hermes-3-Llama-3.1-70BIQ3_S28.79 GiBlegraphista28.79 GiB0.0%
Hermes-3-Llama-3.1-70BIQ3_XS27.29 GiBlegraphista27.29 GiB0.0%
Hermes-3-Llama-3.1-70BIQ2_M22.46 GiBbartowski22.46 GiB0.0%
Hermes-3-Llama-3.1-70BIQ1_S14.29 GiBlegraphista14.29 GiB0.0%
Hermes-3-Llama-3.1-70BIQ3_M29.74 GiBlegraphista29.74 GiB0.0%
Hermes-3-Llama-3.1-70BIQ3_XXS25.58 GiBlegraphista25.58 GiB0.0%
Hermes-3-Llama-3.1-70BIQ2_S20.71 GiBlegraphista20.71 GiB0.0%
Hermes-3-Llama-3.1-70BIQ1_M15.60 GiBbartowski15.60 GiB0.0%
Hermes-3-Llama-3.1-70BIQ1_M15.60 GiBlegraphista15.60 GiB0.0%
Hermes-3-Llama-3.1-70BIQ3_M29.74 GiBbartowski29.74 GiB0.0%
Hermes-3-Llama-3.1-70BQ3_K_S28.79 GiBbartowski28.79 GiB0.0%
Hermes-3-Llama-3.1-70BQ3_K_M31.91 GiBbartowski31.91 GiB0.0%
Hermes-3-Llama-3.1-70BQ3_K_S28.79 GiBlegraphista28.79 GiB0.0%
Hermes-3-Llama-3.1-70BQ3_K_L34.59 GiBbartowski34.59 GiB0.0%
Hermes-3-Llama-3.1-70BQ3_K_L34.59 GiBlegraphista34.59 GiB0.0%

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

ModelQuantizationsSmallest
MN-Violet-Lotus-12B212.79 GiB
Mistral-Small-24B-ArliAI-RPMax-v1.4214.91 GiB
Hermes-3-Llama-3.1-70B1814.29 GiB
Midnight-Miqu-70B-v1.51813.54 GiB
L3-8B-Stheno-v3.3-32K211.88 GiB