Qwen Chat 14B - GGUF

Here are the llama.cpp-compatible GGUF converted and/or quantized models for Qwen 14B Chat.

Explanation of quantization methods

Click to see details

Methods:

type-0 (Q4_0, Q5_0, Q8_0) - weights w are obtained from quants q using w = d * q, where d is the block scale.
type-1 (Q4_1, Q5_1) - weights are given by w = d * q + m, where m is the block minimum

The new methods available are:

GGML_TYPE_Q2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
GGML_TYPE_Q3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This ends up using 3.4375 bpw.
GGML_TYPE_Q4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
GGML_TYPE_Q5_K - "type-1" 5-bit quantization. Same super-block structure as GGML_TYPE_Q4_K resulting in 5.5 bpw
GGML_TYPE_Q6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw
GGML_TYPE_Q8_K - "type-0" 8-bit quantization. Only used for quantizing intermediate results. The difference to the existing Q8_0 is that the block size is 256. All 2-6 bit dot products are implemented for this quantization type.

This is exposed via llama.cpp quantization types that define various "quantization mixes" as follows:

LLAMA_FTYPE_MOSTLY_Q2_K - uses GGML_TYPE_Q4_K for the attention.vw and feed_forward.w2 tensors, GGML_TYPE_Q2_K for the other tensors.
LLAMA_FTYPE_MOSTLY_Q3_K_S - uses GGML_TYPE_Q3_K for all tensors
LLAMA_FTYPE_MOSTLY_Q3_K_M - uses GGML_TYPE_Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K
LLAMA_FTYPE_MOSTLY_Q3_K_L - uses GGML_TYPE_Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K
LLAMA_FTYPE_MOSTLY_Q4_K_S - uses GGML_TYPE_Q4_K for all tensors
LLAMA_FTYPE_MOSTLY_Q4_K_M - uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q4_K
LLAMA_FTYPE_MOSTLY_Q5_K_S - uses GGML_TYPE_Q5_K for all tensors
LLAMA_FTYPE_MOSTLY_Q5_K_M - uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q5_K
LLAMA_FTYPE_MOSTLY_Q6_K- uses 6-bit quantization (GGML_TYPE_Q8_K) for all tensors

Name	Quant method	Bits	Size	Max RAM required	Use case
qwen-chat-14B-Q2_K.gguf	Q2_K	2	6.2 GB	9.1 GB	smallest, significant quality-loss - not recommended for most purposes
qwen-chat-14B-Q3_K_S.gguf	Q3_K_S	3	6.5 GB	9.4 GB	very small, high quality-loss
qwen-chat-14B-Q3_K_M.gguf	Q3_K_M	3	7.2 GB	10.1 GB	very small, high quality-loss
qwen-chat-14B-Q3_K_L.gguf	Q3_K_L	3	7.5 GB	10.4 GB	small, substantial quality-loss
qwen-chat-14B-Q4_0.gguf	Q4_0	4	7.7 GB	10.6 GB	legacy; small, very high quality-loss - prefer using Q3_K_L
qwen-chat-14B-Q4_1.gguf	Q4_1	4	8.4 GB	11.3 GB	legacy; small, very high quality-loss - prefer using Q4_K_S
qwen-chat-14B-Q4_K_S.gguf	Q4_K_S	4	8.0 GB	10.9 GB	small, greater quality-loss
qwen-chat-14B-Q4_K_M.gguf	Q4_K_M	4	8.9 GB	11.8 GB	medium, balanced quality - recommended
qwen-chat-14B-Q5_0.gguf	Q5_0	5	9.2 GB	12.1 GB	legacy; medium, balanced quality - prefer using Q5_K_M
qwen-chat-14B-Q5_1.gguf	Q5_1	5	10 GB	12.9 GB	legacy; medium, balanced quality - prefer using Q5_K_M
qwen-chat-14B-Q5_K_S.gguf	Q5_K_S	5	9.4 GB	12.3 GB	large, low quality-loss - recommended
qwen-chat-14B-Q5_K_M.gguf	Q5_K_M	5	11 GB	13.9 GB	large, very low quality-loss - recommended
qwen-chat-14B-Q6_K.gguf	Q6_K	6	12 GB	14.9 GB	very large, extremely low quality-loss
qwen-chat-14B-Q8_0.gguf	Q8_0	8	15 GB	17.9 GB	very large, extremely low quality-loss - not recommended
qwen-chat-14B-f16.gguf	f16	16	27 GB	29.9 GB	very large, no quality-loss - not recommended