bartowski/Intern-S2-397B-GGUF

Llamacpp imatrix Quantizations of Intern-S2-397B by internlm

Using llama.cpp release b10896 for quantization.

Original model: https://huggingface.co/internlm/Intern-S2-397B

Model details:

• Parameter count: 403B
• Input support: text, image (with mmproj file) - details
• Speculative decoding: yes (MTP) - details
• imatrix: yes - details

How to run

Prompt format

No prompt format found, check original model page

Don't know which to choose? Grab Q4_K_M (254.58GB) - usually a good mix of size and performance. Download instructions available here

Available files:

Filename	Quant type	File Size	Split	Description
Intern-S2-397B-Q8_0.gguf	Q8_0	428.62GB	true	Extremely high quality, generally unneeded but max available quant.
Intern-S2-397B-Q6_K.gguf	Q6_K	353.55GB	true	Very high quality, near perfect, recommended.
Intern-S2-397B-Q6_K_S.gguf	Q6_K_S	336.90GB	true	Very high quality, near perfect, a little smaller than Q6_K with almost all of the model at Q6_K precision, recommended.
Intern-S2-397B-Q5_K_M.gguf	Q5_K_M	309.36GB	true	High quality, recommended.
Intern-S2-397B-Q5_K_S.gguf	Q5_K_S	286.57GB	true	High quality, recommended.
Intern-S2-397B-Q4_K_M.gguf	Q4_K_M	254.58GB	true	Good quality, default size for most use cases, recommended.
Intern-S2-397B-IQ4_NL.gguf	IQ4_NL	253.29GB	true	Similar to IQ4_XS, but slightly larger.
Intern-S2-397B-Q4_1.gguf	Q4_1	252.83GB	true	Legacy format, similar performance to Q4_K_S but with improved tokens/watt on Apple silicon.
Intern-S2-397B-Q4_K_S.gguf	Q4_K_S	237.23GB	true	Slightly lower quality with more space savings, recommended.
Intern-S2-397B-Q4_0.gguf	Q4_0	229.17GB	true	Legacy format, kept for compatibility with older tools.
Intern-S2-397B-IQ4_XS.gguf	IQ4_XS	223.27GB	true	Decent quality, smaller than Q4_K_S with similar performance, recommended.
Intern-S2-397B-IQ3_M.gguf	IQ3_M	213.19GB	true	Medium-low quality, new method with decent performance comparable to Q3_K_M.
Intern-S2-397B-Q3_K_L.gguf	Q3_K_L	203.95GB	true	Lower quality but usable, good for low RAM availability.
Intern-S2-397B-Q3_K_M.gguf	Q3_K_M	193.69GB	true	Low quality.
Intern-S2-397B-Q3_K_S.gguf	Q3_K_S	183.78GB	true	Low quality, not recommended.
Intern-S2-397B-IQ3_XS.gguf	IQ3_XS	181.84GB	true	Lower quality, new method with decent performance, slightly better than Q3_K_S.
Intern-S2-397B-IQ3_XXS.gguf	IQ3_XXS	173.29GB	true	Lower quality, new method with decent performance, comparable to Q3 quants.
Intern-S2-397B-Q2_K.gguf	Q2_K	154.50GB	true	Very low quality but surprisingly usable.
Intern-S2-397B-IQ2_M.gguf	IQ2_M	146.06GB	true	Relatively low quality, uses SOTA techniques to be surprisingly usable.
Intern-S2-397B-IQ2_S.gguf	IQ2_S	131.56GB	true	Low quality, uses SOTA techniques to be usable.
Intern-S2-397B-IQ2_XS.gguf	IQ2_XS	125.16GB	true	Low quality, uses SOTA techniques to be usable.
Intern-S2-397B-IQ2_XXS.gguf	IQ2_XXS	118.84GB	true	Very low quality, uses SOTA techniques to be usable.
Intern-S2-397B-IQ1_M.gguf	IQ1_M	101.82GB	true	Extremely low quality, not recommended.
Intern-S2-397B-IQ1_S.gguf	IQ1_S	91.61GB	true	Extremely low quality, not recommended.

Download a specific file:

hf download bartowski/Intern-S2-397B-GGUF --include "Intern-S2-397B-Q4_K_M/*" --local-dir ./

Downloading using the Hugging Face CLI

Click to view download instructions

First, make sure you have the Hugging Face CLI installed:

pip install -U "huggingface_hub[cli]"

Download a specific file:

hf download bartowski/Intern-S2-397B-GGUF --include "Intern-S2-397B-Q4_K_M/*" --local-dir ./

The files marked true in the Split column above are stored as multiple parts in a folder. To download all the parts to a local folder, run:

hf download bartowski/Intern-S2-397B-GGUF --include "Intern-S2-397B-Q8_0/*" --local-dir ./

You can either specify a new local-dir (Intern-S2-397B-Q8_0) or download them all in place (./)

How to run

These quants run with llama.cpp - installable in one line via llama.app:

curl -LsSf https://llama.app/install.sh | sh
llama-server -hf bartowski/Intern-S2-397B-GGUF:Q4_K_M

llama-server includes a built-in chat web UI, served at http://localhost:8080 by default.

These quants were made with llama.cpp release b10896 - if this model's architecture is newly supported, you'll need that release or newer to run them.

They also work in: LM Studio · koboldcpp · ramalama · Jan AI · Text Generation Web UI · LoLLMs · Atomic Chat

Multimodal

This model supports image input. Alongside the quants, this repo includes the multimodal projector files mmproj-Intern-S2-397B-f16.gguf and mmproj-Intern-S2-397B-bf16.gguf, which pair with any quant above.

llama.cpp downloads the mmproj automatically when using -hf as shown above; if you're loading files manually, pass it with --mmproj.

MTP

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 the imatrix quants (except for the Q8_0 quant), since imatrix calibration does not exercise them. Q4_0 is chosen for its speed which massively benefits MTP performance.

Per-tensor layouts

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-61746c35382fd0e6 is the exact snapshot these files record. The method is described in this write-up.

Checked on bartowski/Nex-N2.5-Pro-GGUF (same architecture and tensor shapes) before any of these files were released: Q4_K_M reached 0.83×, Q3_K_M 0.76× and IQ1_S 0.80× 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	353.55GB	7.04	7.01	70 %
Q6_K_S	336.90GB	6.71	6.68	91 %
Q5_K_M	309.36GB	6.15	6.13	70 %
Q5_K_S	286.57GB	5.68	5.68	90 %
Q4_K_M	254.58GB	5.04	5.05	70 %
IQ4_NL	253.29GB	5.01	5.02	70 %
Q4_K_S	237.23GB	4.68	4.70	90 %
IQ4_XS	223.27GB	4.40	4.43	90 %
IQ3_M	213.19GB	4.20	4.23	50 %
Q3_K_L	203.95GB	4.01	4.04	50 %
Q3_K_M	193.69GB	3.80	3.84	70 %
Q3_K_S	183.78GB	3.60	3.64	90 %
IQ3_XS	181.84GB	3.56	3.61	90 %
IQ3_XXS	173.29GB	3.38	3.44	70 %
Q2_K	154.50GB	3.00	3.06	70 %
IQ2_M	146.06GB	2.83	2.90	70 %
IQ2_S	131.56GB	2.54	2.61	70 %
IQ2_XS	125.16GB	2.41	2.48	90 %
IQ2_XXS	118.84GB	2.28	2.36	70 %
IQ1_M	101.82GB	1.93	2.02	70 %
IQ1_S	91.61GB	1.73	1.82	70 %

Checked on bartowski/Nex-N2.5-Pro-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	0.0285 ± 0.0005	0.0396 ± 0.0007	0.83×	+3.7 %
Q3_K_M	0.0700 ± 0.0012	0.1063 ± 0.0017	0.76×	+4.7 %
IQ1_S	0.5358 ± 0.0063	0.7574 ± 0.0077	0.80×	+7.4 %

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.

imatrix

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: Intern-S2-397B-calibration-v6.txt. The imatrix is available here: Intern-S2-397B-imatrix.gguf.

Calibration render details

{
  "generator": "auto_quant_v2 calibration renderer",
  "recipe": "calibration-v6",
  "model": "Intern-S2-397B",
  "encoder": "chat_template",
  "chunk_size": 512,
  "prose_chunks": 214,
  "tool_chunks": 335,
  "total_chunks": 549,
  "tool_chunk_fraction": 0.61,
  "n_conversations": 137,
  "extension_convs_used": 0,
  "conversation_token_lengths": [
    565,
    1628,
    915,
    1403,
    1146,
    1365,
    2567,
    665,
    1228,
    1233,
    1068,
    2097,
    882,
    977,
    2401,
    1274,
    1121,
    987,
    738,
    719,
    1363,
    1059,
    1408,
    1203,
    1719,
    1499,
    1051,
    723,
    973,
    1504,
    1570,
    841,
    1269,
    1038,
    1017,
    1713,
    1652,
    1176,
    478,
    1911,
    1412,
    1127,
    1399,
    1771,
    1962,
    1316,
    1610,
    782,
    2442,
    1103,
    2638,
    788,
    1025,
    752,
    904,
    699,
    2314,
    759,
    1149,
    1088,
    1232,
    1172,
    927,
    1023,
    969,
    1424,
    914,
    1553,
    1836,
    842,
    322,
    1123,
    3002,
    2837,
    710,
    763,
    1033,
    796,
    1299,
    1083,
    1152,
    796,
    1072,
    1072,
    1292,
    1547,
    1404,
    1390,
    874,
    659,
    2471,
    639,
    1162,
    1676,
    1948,
    1204,
    644,
    1357,
    1125,
    1730,
    1699,
    1716,
    818,
    771,
    1038,
    2857,
    741,
    597,
    764,
    1262,
    1053,
    1371,
    769,
    349,
    318,
    2322,
    734,
    1140,
    1831,
    1729,
    2364,
    2374,
    692,
    981,
    841,
    851,
    1240,
    1000,
    843,
    1358,
    831,
    711,
    1422,
    1028,
    918,
    1317,
    1480
  ],
  "warnings": []
}

ARM/AVX information

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.

Which file should I choose?

Click here for details

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:

llama.cpp feature matrix

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.

Credits

Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset.

Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski