GGUF Q3_K_L (3-bit K-quant, large)
4.1200 measured avg (llama.cpp quantize README, Llama-3.1-8B). Pure block_q3_K is 3.4375 bpw; Q3_K_L is a mixture promoting attention and feed-forward tensors to Q5_K.
Fields
id
quantization/gguf-q3-k-lScheme
gguf-q3_k_m
Bits
4.1200 measured avg (llama.cpp quantize README, Llama-3.1-8B). Pure block_q3_K is 3.4375 bpw; Q3_K_L is a mixture promoting attention and feed-forward tensors to Q5_K.
Weight group size
16
Kernels
ggml dequantize_row_q3_K + get_rowsggml-cuda mul_mat_vec_q / dequantize_mul_mat_vec + convert_mul_mat_q
Quality delta
Llama-3.1-8B-Instruct vs FP16 baseline from Table 2 of Kurt (2026, https://arxiv.org/abs/2601.14277, quant-gguf-kquant-study): WikiText-2 perplexity 7.81 vs 7.32 (+0.49 ppl delta); downstream 5-benchmark average 68.78 vs 69.47 (-0.69 delta): GSM8K 74.07 vs 77.63 (-3.56 pp), HellaSwag 73.54 vs 72.51 (+1.03 pp), IFEval 79.14 vs 78.93 (+0.21 pp), MMLU 62.31 vs 63.50 (-1.19 pp), TruthfulQA MC2 54.84 vs 54.79 (+0.05 pp). Model size reduction 73.14% vs FP16.
notes
Q3_K_L is the large variant of the 3-bit K-quant family in llama.cpp. Base block structure is identical to Q3_K (16-value scale group, 16 6-bit sub-scales per 256 super-block, 110 bytes / 256 = 3.4375 bpw), but the quantization recipe (in src/llama-quant.cpp) promotes more critical tensors (such as attention.wv and feed-forward down-projections) to Q5_K compared to Q3_K_M. In Kurt (2026), Q3_K_L preserves downstream task performance substantially better than Q3_K_S (68.78 avg vs 65.49 avg) while achieving 73.14% size reduction compared to FP16. SCHEMA PROXY: scheme uses 'gguf-q3_k_m' because schemas/quantization.schema.json has no separate 'gguf-q3_k_l' enum value.
Notes
Q3_K_L is the large variant of the 3-bit K-quant family in llama.cpp. Base block structure is identical to Q3_K (16-value scale group, 16 6-bit sub-scales per 256 super-block, 110 bytes / 256 = 3.4375 bpw), but the quantization recipe (in src/llama-quant.cpp) promotes more critical tensors (such as attention.wv and feed-forward down-projections) to Q5_K compared to Q3_K_M. In Kurt (2026), Q3_K_L preserves downstream task performance substantially better than Q3_K_S (68.78 avg vs 65.49 avg) while achieving 73.14% size reduction compared to FP16. SCHEMA PROXY: scheme uses 'gguf-q3_k_m' because schemas/quantization.schema.json has no separate 'gguf-q3_k_l' enum value.