Everything behind the suffixes on Hugging Face: what quantization actually does, the full GGUF alphabet (Q4_K_M, Q6_K, IQ quants, UD, imatrix), the other formats (AWQ, GPTQ, EXL3, MLX, FP8, NVFP4, QAT), the fine-tune zoo (instruct, thinking, uncensored, flash, MTP), how to decode any repo, and a step-by-step framework for choosing the right model for your hardware — using Qwen3.8-27B as the running example.
Precision, bit math, and why a 54 GB model can run on a 24 GB GPU.
Bits, groups, scales, bpw, and why some weights matter more than others.
Q4_K_M, Q5_K_S, Q6_K, Q8_0, XL tiers and the legacy Q4_0 — with real Qwen3.8-27B file sizes.
IQ1/IQ2/IQ3/IQ4, importance matrices, and when going under 4 bits is a good idea.
AWQ, GPTQ, EXL2/EXL3, MLX, bitsandbytes NF4, FP8, NVFP4, QAT, AQLM, HQQ, EETQ, Marlin.
instruct vs base, thinking, uncensored/abliterated, flash, vision, MTP, MoE, dates, merges and more.
Filename anatomy, what to verify, trusted publishers, and red flags.
The decision framework: hardware → budget → model size → quant → ecosystem → test.
16-question comprehensive quiz across all lessons.