GGUF vs GPTQ vs AWQ vs EXL2: LLM Model Formats Explained (2026)

GGUF vs GPTQ vs AWQ vs EXL2: LLM Model Formats Explained (2026)

First, separate 2 ideas: containers vs. quantization methods

Most confusion comes from mixing 2 layers.

A container defines how tensors are stored on disk. A quantization method defines how weights are squeezed into fewer bits.

Containers: safetensors, GGUF, PyTorch pickle (.bin / .pt).

Methods: GPTQ, AWQ, bitsandbytes NF4, llama.cpp K-quants and I-quants.

Both at once: EXL2 and EXL3 are a method plus a storage layout tied to one inference library.

Weight memory ≈ parameters × bits-per-weight ÷ 8.

Model16-bit~4.5 bits per weight8B~16 GB~4.5 GB70B~140 GB~39 GB

This is arithmetic, not a vendor benchmark. It covers weights only. The KV cache and runtime overhead add more on top.

1. Full precision: safetensors and PyTorch .bin

Unquantized models usually ship as 16-bit weights, in either pytorch_model.bin or model.safetensors.

The older .bin / .pt files use Python pickle. Loading a pickle file can execute arbitrary code, which makes untrusted checkpoints a security risk.

Safetensors, created at Hugging Face, removes that risk. A file is a small JSON header plus raw tensor buffers, with nothing executable inside. Tensors can be memory-mapped and loaded one at a time without reading the whole file. Safetensors is now listed as a PyTorch Foundation project.

Important nuance: most GPTQ, AWQ, EXL2, EXL3, and MLX models are also stored in .safetensors files. The quantization lives in the tensor contents and a config file, not in a new container.

GGUF is a binary format for running models with GGML and GGML-based executors such as llama.cpp. It was created by Georgi Gerganov, who also leads llama.cpp (Hugging Face docs). It was introduced on August 21, 2023 as the replacement for the older GGML format.

The older GGML, GGMF, and GGJT files could not say which architecture a model belonged to. Adding a new hyperparameter broke every existing file. GGUF switched to typed key-value metadata, so new fields can be added without breaking old files.

The spec lists 5 goals: single-file deployment, extensibility, mmap compatibility, easy loading, and complete information inside the file. Unlike tensor-only formats, GGUF can carry the tokenizer, special tokens, and a Jinja chat template alongside the weights.

The suffix in a name like Q4_K_M.gguf tells you the scheme. Figures below come from the Hugging Face GGUF docs.

TypeHow it worksBits per weightQ4_0 / Q4_1 (legacy)4-bit round-to-nearest in 32-weight blocks; Q4_1 adds a block minimum4.5 / 5.0*Q8_0 (legacy label)8-bit round-to-nearest in 32-weight blocks8.5*Q2_K16 blocks × 16 weights per super-block, 4-bit scales and mins2.625Q3_K16 blocks × 16 weights, 6-bit scales3.4375Q4_K8 blocks × 32 weights, 6-bit scales and mins4.5Q5_K8 blocks × 32 weights, 6-bit scales and mins5.5Q6_K16 blocks × 16 weights, 8-bit scales6.5625IQ4_XS256-weight super-blocks, uses an importance matrix4.25IQ3_XXSSame I-quant family3.06IQ2_XXSSame I-quant family2.06IQ1_SSame I-quant family1.56

*Derived by hand, not listed in the HF table: 32 weights plus a 16-bit scale (and a 16-bit minimum for Q4_1).

Checking the Q4_K math: A super-block holds 256 weights. 256 × 4 bits = 1,024 bits. Add 8 blocks × 12 bits of scales and minimums (96 bits). Add a 16-bit super-scale and 16-bit super-minimum (32 bits). Total: 1,152 ÷ 256 = 4.5 bits per weight.

What _S, _M, _L mean: These are mixes, not new types. For example, llama.cpp describes Q4_K_M as using Q6_K for half of the attention.wv and feed_forward.w2 tensors and Q4_K elsewhere (Unsloth docs). That is why a Q4_K_M file averages above 4.5 bits per weight.

Newer types: The HF table also lists TQ1_0 and TQ2_0 for ternary weights, plus MXFP4, a 4-bit microscaling floating-point type.

A labeling quirk: Hugging Face files Q8_0 under legacy types. In practice, Q8_0 remains the standard near-lossless GGUF choice.

Hugging Faces reference table for a Llama-2-7B-class model shows the trade-off:

QuantPerplexityChange vs FP16SizeFP165.9565baseline13.0 GBQ8_05.9584+0.03%7.0 GBQ6_K5.9642+0.13%5.5 GBQ5_K_M5.9796+0.39%4.8 GBQ4_K_M6.0565+1.68%4.1 GB

Illustrative only. These numbers come from a 2023-era 7B model; newer models can react differently.

GGUF quantization can use calibration data. llama.cpps llama-imatrix computes an importance matrix from a text file. llama-quantize --imatrix then uses it to improve quality. For 1-bit and 2-bit mixes, llama-quantize warns if no imatrix is supplied.

The spec defines filenames as base name, size label, fine-tune, version, encoding, type, and shard. Shards use a 5-digit counter such as 00003-of-00009. Optional mmproj- and mtp- prefixes mark vision projectors and multi-token-prediction draft modules.

GGUF is native to llama.cpp and its ecosystem. Hugging Face documents use with llama.cpp, LM Studio, GPT4All, and Ollama.

vLLM support exists but is limited. vLLM calls it highly experimental and under-optimized, and GGUF now needs the out-of-tree vllm-gguf-plugin.

GPTQ was written by Elias Frantar (IST Austria), Saleh Ashkboos and Torsten Hoefler (ETH Zurich), and Dan Alistarh (IST Austria & Neural Magic). It first appeared on arXiv on October 31, 2022. It was published at ICLR 2023.

GPTQ is a one-shot, post-training weight quantization method. It uses approximate second-order (Hessian) information to decide how to round weights. Rounding error in one column is compensated by adjusting weights not yet quantized. It needs a small calibration dataset but no retraining.

Quantized 175B-parameter models in about 4 GPU hours, down to 3 or 4 bits per weight (arXiv).

Reported negligible accuracy loss at those bit widths.

End-to-end speedups over FP16 of about 3.25x on NVIDIA A100 and 4.5x on A6000 (HF paper page).

GPTQ repos often include GPTQ or tags like 4bit-128g in the name.

Group size (128g): one scale per 128 weights. Smaller groups improve accuracy but add a little size.

Act-order (desc_act): quantizes columns in order of importance, usually improving accuracy.

The original AutoGPTQ library is no longer maintained.

GPTQModel states it has fully supplanted AutoGPTQ and AutoAWQ for Transformers, Optimum, and PEFT. Its output runs in Transformers, vLLM, and SGLang.

llm-compressor also implements GPTQ, but saves results in the compressed-tensors format.

Hugging Face estimates GPTQ calibration for an 8B model at about 20 minutes on 1 A100.

AWQ (Activation-aware Weight Quantization) comes from Song Hans group at MIT. It first appeared on arXiv on June 1, 2023. It won the MLSys 2024 Best Paper Award.

Not all weights matter equally. Protecting roughly 1% of salient weights sharply reduces quantization error.

The twist: AWQ finds those salient channels by looking at activation magnitudes, not the weights themselves.

It does not store those channels at higher precision. Instead, it scales them up through a mathematically equivalent transformation, keeping a uniform, hardware-friendly format. AWQ uses no backpropagation or reconstruction, so it is less likely to overfit its calibration set.

The papers TinyChat runtime ran more than 3x faster than the Hugging Face FP16 implementation on desktop and mobile GPUs.

Hugging Face estimates AWQ calibration for an 8B model at about 10 minutes on 1 A100, roughly half of GPTQs estimate.

AutoAWQ is officially deprecated. Its last tested setup was Torch 2.6.0 and Transformers 4.51.3.

vLLM adopted the functionality into llm-compressor, now the recommended AWQ workflow.

MLX-LM also supports AWQ on Apple Silicon.

EXL2 is the native format of ExLlamaV2, an inference library by turboderp for consumer GPUs. It uses the same optimization method as GPTQ and supports 2, 3, 4, 5, 6, and 8-bit quantization.

Any average bitrate from 2 to 8 bits per weight: Quantization levels can be mixed across and within layers.

Column-level mixing: More important columns inside a layer can get more bits.

Automatic allocation: The converter quantizes each matrix several