Quantization publisher

legraphista

legraphista publishes 1,122 quantizations across 47 models in our index, averaging 4.563 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
48
Quantizations
1,122
Models covered
47
Avg effective bpw
4.563
across their files

Same model, same quant label, different bytes

largest disagreements first
ModelQuantlegraphistavsTheirsDifference
c4ai-command-r-plus-08-2024Q8_080.54 GiBbartowski102.74 GiB-21.6%
Llama-Guard-3-8BQ8_07.95 GiBeaddario7.20 GiB+10.4%
Llama-Guard-3-8BQ3_K_L4.03 GiBeaddario3.50 GiB+15.0%
Llama-Guard-3-8BQ4_K_S4.37 GiBeaddario3.99 GiB+9.7%
Llama-Guard-3-8BQ3_K_S3.41 GiBeaddario3.08 GiB+10.9%
Llama-Guard-3-8BQ5_K_S5.21 GiBeaddario4.88 GiB+6.8%
Llama-Guard-3-8BIQ4_NL4.36 GiBeaddario4.09 GiB+6.4%
Llama-Guard-3-8BIQ3_S3.43 GiBeaddario3.20 GiB+7.3%
shieldgemma-27bIQ4_XS13.80 GiBmradermacher13.92 GiB-0.9%
Llama-Guard-3-8BIQ3_M3.52 GiBeaddario3.43 GiB+2.7%
DeepSeek-Coder-V2-Lite-InstructIQ4_XS7.98 GiBmradermacher8.05 GiB-0.9%
codegeex4-all-9bIQ4_XS4.89 GiBmradermacher4.94 GiB-1.0%
glm-4-9b-chatIQ4_XS4.94 GiBbartowski4.89 GiB+1.0%
glm-4-9b-chat-1mIQ4_XS4.98 GiBbartowski4.93 GiB+1.0%
Llama3-ChatQA-1.5-8BIQ4_XS4.14 GiBPrunaAI4.18 GiB-0.8%
Llama-Guard-3-8BIQ4_XS4.14 GiBmradermacher4.18 GiB-0.8%
Mistral-7B-v0.3IQ4_XS3.64 GiBmradermacher3.68 GiB-0.9%
Llama3-ChatQA-1.5-8BIQ4_NL4.36 GiBPrunaAI4.38 GiB-0.6%
Llama-Guard-3-8BQ6_K6.14 GiBeaddario6.12 GiB+0.4%
Llama-3.1-Minitron-4B-Width-BaseIQ4_XS2.38 GiBbartowski2.36 GiB+0.7%
Llama3-ChatQA-1.5-8BQ5_K_S5.21 GiBQuantFactory5.22 GiB-0.0%
Llama3-ChatQA-1.5-8BQ6_K6.14 GiBQuantFactory6.14 GiB-0.0%
Llama3-ChatQA-1.5-8BQ8_07.95 GiBQuantFactory7.95 GiB-0.0%
Llama3-ChatQA-1.5-8BQ3_K_L4.03 GiBQuantFactory4.03 GiB-0.0%
Llama3-ChatQA-1.5-8BQ3_K_S3.41 GiBQuantFactory3.41 GiB-0.0%

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