phi-4 - DASH-Q 2-bit GGUF

llama.cpp GGUF files of microsoft/phi-4 quantized with DASH-Q, at several 2-bit-class sizes. Every file uses only standard llama.cpp tensor types (no tensor above 4 bits) and loads in any recent llama.cpp build.

File	Type	Size	Bits / weight
`phi-4-DASHQ-IQ2_XXS.gguf`	IQ2_XXS	4.66 GB	2.54
`phi-4-DASHQ-IQ2_XS.gguf`	IQ2_XS	5.18 GB	2.83
`phi-4-DASHQ-IQ2_M.gguf`	IQ2_M	5.61 GB	3.07
`phi-4-DASHQ-Q2_K_XL.gguf`	Q2_K_XL	6.06 GB	3.31

Perplexity (lower is better)

Type	Model	Size	WikiText-2	C4
	F16	29.3 GB	5.78	11.27
IQ2_XXS	llama.cpp IQ2_XXS (imatrix)	4.55 GB	7.59	13.94
IQ2_XXS	DASH-Q IQ2_XXS	4.66 GB	7.38	13.80
IQ2_XS	llama.cpp IQ2_XS (imatrix)	4.92 GB	7.01	13.05
IQ2_XS	DASH-Q IQ2_XS	5.18 GB	6.56	12.47
IQ2_M	llama.cpp IQ2_M (imatrix)	5.49 GB	6.46	12.28
IQ2_M	DASH-Q IQ2_M	5.61 GB	6.27	11.97
Q2_K_XL	llama.cpp Q2_K (imatrix)	5.92 GB	6.42	12.17
Q2_K_XL	unsloth Q2_K_L	5.73 GB	6.74	12.61
Q2_K_XL	DASH-Q Q2_K_XL	6.06 GB	6.18	11.87

llama-perplexity, context 2048; WikiText-2 test, C4 validation (256 x 2048 tokens).

Zero-shot accuracy (lm-eval-harness, 0-shot)

File	arc_easy	arc_challenge	piqa	hellaswag	winogrande	openbookqa	commonsense_qa	truthfulqa_mc2	lambada_openai	avg
DASH-Q IQ2_XXS	77.5	53.1	79.8	75.0	75.1	43.8	72.7	54.2	75.8	67.44
DASH-Q IQ2_XS	76.3	56.1	81.0	78.4	76.4	46.2	76.0	55.9	74.7	69.01
DASH-Q IQ2_M	74.2	54.8	81.2	79.2	74.0	44.6	76.8	56.7	75.5	68.56
DASH-Q Q2_K_XL	74.4	55.5	81.4	80.2	76.9	44.6	76.2	56.4	74.2	68.86

Usage

llama-cli -m phi-4-DASHQ-IQ2_M.gguf -ngl 99 -c 8192

License

Inherits the license of the base model (microsoft/phi-4).