<|im_start|>system
{system_prompt}
# Tools
You may call one or more functions to assist with the user query.
You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type": "function", "function": {"name": "get_stock_price", "description": "Get the current stock price", "parameters": {"type": "object", "properties": {"symbol": {"type": "string", "description": "The stock symbol, e.g. AAPL, GOOG"}}, "required": ["symbol"]}}}
</tools>
For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call><|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
Don't know which to choose? Grab Q4_K_M (9.01GB) - usually a good mix of size and performance. Download instructions available here
Relatively low quality, uses SOTA techniques to be surprisingly usable.
About these sizes: 30.5 % of this model's weights (ffn_down_exps (896 wide)) have row widths that llama.cpp's 256-wide K and IQ block types cannot use, so each size gives them one of the 32-wide types instead (Q4_0 to Q8_0, or IQ4_NL), whatever type its name asks for, and the files above are closer together in size than usual.
This model has MTP (Multi-Token Prediction) layers, and they are included in these quants
MTP layers act as a built-in draft model, letting llama.cpp run speculative decoding for faster generation. To use them, add the following flag to your llama.cpp command:
--spec-type draft-mtp
Note: the MTP layers are stored at Q4_0 in every quant except Q8_0. Imatrix calibration does not exercise them, so the imatrix has nothing to offer these layers, and Q4_0 is chosen for its speed, which massively benefits MTP performance.
Some of these files were built with a layout computed for this model instead of llama.cpp's standard one-size-fits-all rules. A Q4_K_M is still mostly Q4_K; the extra precision goes to the weights this particular model is most sensitive to. The S, M or L in a name says how much of the model stays at the base precision: about 90 % for S, 70 % for M and 50 % for L. An _L name is simply the large size of its family. Q4_K_L is to Q4_K_M what Q4_K_M is to Q4_K_S; Q6_K_S, Q6_K and Q6_K_L are the small, medium and large sizes of Q6_K, with Q6_K_L about halfway to Q8_0. In earlier releases an _L name meant the embedding and output weights were kept at Q8_0; in these files it means the larger size of the base type. There is no size target, so each file's bits per weight is reported rather than promised.
The layout each of these files was built with is published in the layouts/ folder: <file>.tensor-types.txt is the exact --tensor-type-file given to llama-quantize, and <file>.layout.json records how it was computed, including the generator version, the llama.cpp release and the commit, so any of them can be rebuilt. The code that computed them is public at quantization-config; its tag key-3bf8b43e20a68f0b is the exact snapshot these files record. The method is described in this write-up.
Checked on bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF (same architecture and tensor shapes) before any of these files were released: Q4_K_M reached 0.91×, Q3_K_M 0.92× and IQ2_M 1.09× the KL divergence of the standard layout at the same file size.
Layout details
Files built from a computed layout:
Quant
Size
Body bits/weight
File bits/weight
Body kept at base type
Q6_K_L
12.05GB
7.38
7.66
51 %
Q6_K
11.70GB
7.03
7.44
71 %
Q6_K_S
11.38GB
6.71
7.23
91 %
Q5_K_M
10.17GB
6.12
6.47
71 %
Q4_K_L
9.42GB
5.49
5.99
50 %
Q5_K_S
9.20GB
5.69
5.85
90 %
Q4_K_M
9.01GB
5.08
5.73
70 %
Q4_K_S
8.09GB
4.70
5.15
90 %
IQ4_NL
7.90GB
5.08
5.03
70 %
IQ3_M
7.46GB
4.32
4.74
50 %
Q3_K_L
7.35GB
4.20
4.67
50 %
IQ4_XS
7.21GB
4.39
4.59
90 %
Q3_K_M
7.01GB
3.98
4.46
70 %
IQ3_XS
6.60GB
3.57
4.20
90 %
IQ3_XXS
6.51GB
3.47
4.14
70 %
Q3_K_S
6.29GB
3.56
4.00
90 %
Q2_K
5.77GB
3.04
3.67
70 %
IQ2_M
5.73GB
2.90
3.64
70 %
Checked on bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF, which shares this model's architecture and tensor shapes: the computed layout against the standard one, measured by KL divergence against the unquantized model. The ratio compares each computed file with the standard ladder read at that file's own size, so it can differ from the two KLD columns when the two files differ in size.
Quant
Computed layout KLD
Standard layout KLD
Ratio at equal size
Size vs standard file
Q4_K_M
0.0321 ± 0.0004
0.0509 ± 0.0006
0.91×
+8.0 %
Q3_K_M
0.1059 ± 0.0013
0.1537 ± 0.0017
0.92×
+8.5 %
IQ2_M
0.2690 ± 0.0028
0.2983 ± 0.0031
1.09×
+6.0 %
How it works: the base type is a floor for every body tensor and a fixed share of the body bytes stays at it (90 % for S, 70 % for M, 50 % for L); the remaining bytes go where a cross-model sensitivity prior, measured by KL divergence against the unquantized model, says they buy the most quality. The embedding and output tensors are sized by their share of the file: a small table is kept at Q8_0, a large one follows the file's bitrate. A K-quant and the IQ quant with the same base bitrate (Q3_K_S and IQ3_XS, Q3_K_M and IQ3_S, Q3_K_L and IQ3_M) come out at about the same size; the IQ file is the GPU-oriented twin.
All quants made using imatrix option, with a calibration corpus rendered through this model's own chat template. The corpus pairs plain prose with tool-calling and reasoning conversations (corpus source data), encoded exactly as this model sees them at inference and processed with --parse-special, so chat-format special tokens contribute to the importance matrix. The corpus rendered for this model is included in this repo: Mellum2.1-12B-A2.5B-Thinking-MTP-calibration-v6.txt. The imatrix is available here: Mellum2.1-12B-A2.5B-Thinking-MTP-imatrix.gguf.
llama.cpp automatically "repacks" weights into an interleaved layout at load time for faster inference on ARM and AVX machines - details in this PR. This once required downloading special Q4_0_4_4/4_8/8_8 files; those are long gone. Online repacking now covers Q4_0, IQ4_NL, and most K-quants, so no special quant choice is needed for CPU inference.
An older (early 2024) but still useful write-up with charts comparing quant performances is provided by Artefact2 here
The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.
If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.
If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.
Hugging Face can also do this math for you: add your hardware in your Local Apps settings and the model page will show which files fit.
Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.
If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.
If you want to get more into the weeds, you can check out this extremely useful feature chart:
But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.
These I-quants can also be used on CPU, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.
