Mixed Precision GGUF layer quantization of Llama-4-Scout-17B-16E-Instruct by meta-llama

Original model: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct

The hybrid quant employs different quantization levels on a per layer basis to enable both high performance and small file size at the same time. An extended layer definition E quant Q4_E_H for the model is defined as follows (updated 10/27/2026):

   LAYER_TYPES='[
   ["A","attn","Q","attn_q","K","attn_k","V","attn_v","O","attn_o","S","ssm","F","ffn","G","ffn_g","U","ffn_u","D","ffn_d"],
   ["MAP","VOFD","0","QN_K","2","Q2_K","3","Q3_K","4","Q4_K","5","Q5_K","6","Q6_K","8","Q8_0","h","F16","f","F32"],
   [0 ,"Q5_K_6545"],[1 ,"Q5_K_5544"],[2 ,"Q4_K_5544"],[3 ,"Q4_K_5544"],[4 ,"Q4_K_5544"],[5 ,"Q4_K_5544"],[6 ,"Q4_K_5544"],[7 ,"Q4_K_5544"],
   [8 ,"Q4_K_5544"],[9 ,"Q4_K_5544"],[10,"Q4_K_5544"],[11,"Q4_K_5544"],[12,"Q4_K_5544"],[13,"Q4_K_5544"],[14,"Q4_K_5544"],[15,"Q4_K_5544"],
   [16,"Q4_K_5544"],[17,"Q4_K_5544"],[18,"Q4_K_5544"],[19,"Q4_K_5544"],[20,"Q4_K_5544"],[21,"Q4_K_5544"],[22,"Q4_K_5544"],[23,"Q4_K_5544"],
   [24,"Q4_K_5544"],[25,"Q4_K_5544"],[26,"Q4_K_5544"],[27,"Q4_K_5544"],[28,"Q4_K_5544"],[29,"Q4_K_5544"],[30,"Q4_K_5544"],[31,"Q4_K_5544"],
   [32,"Q4_K_5544"],[33,"Q4_K_5544"],[34,"Q4_K_5544"],[35,"Q4_K_5544"],[36,"Q4_K_5544"],[37,"Q4_K_5544"],[38,"Q4_K_5544"],[39,"Q4_K_5544"],
   [40,"Q4_K_5544"],[41,"Q4_K_5544"],[42,"Q4_K_5544"],[43,"Q4_K_5544"],[44,"Q4_K_5544"],[45,"Q4_K_5545"],[46,"Q5_K_6656"],[47,"Q6_K_8886"]
   ]'
   FLAGS="--token-embedding-type Q6_K --output-tensor-type Q6_K"

This quant was sized to enable operation on a 64G ram machine while using minimum Q4_K quant level across layers to minimize entropy loss. Layers were tuned for good performance across a small set of reasoning prompts.

LEGACY QUANTS:

Older smaller legacy quants remain available since the model is still pulling appreciable downloads per month they are left on the model page. The new E quant is recommended for use if hardware platform can support running it.

Q4_K_H quant achieves a ~50G gguf with the same perplexity as a ~60G IQ4_XS gguf. The quants employed are all K to avoid slow CPU processing of IQ quants, and they range from Q3_K_S at early layers to a final layer which may be specified by the quant name. In this case the final layer is Q4_K_M so the quant is called Q4_K_H. Note there is no unique Q4_K_H quant since the selections of quantizations to use as a function of layer are arbitrary. For this file the layer quants are as follows:

embed  : Q3_K
0..7   : Q3_K_S
8..23  : alt Q3_K_M Q3_K_S
24..44 : Q3_K_M
45     : Q3_K_L
46     : Q4_K_S
47     : Q4_K_M
output : Q5_K

These quants were select based on combined subjective and objective performance evaluations to give both high performance and moderate file size.

A slightly smaller version of this model is available which intentionally drops to Q2_K on alternate deep layers to add a small controlled amount of extra nonlinearity and possibly help boostrap better generation creativity for vision mode. It is named Q4_V_H (V=Vision, but it does not have to be used for vision it will work for text only input also) to distinguish it from Q4_K_H. This quant has the side benefit of just fitting under 50G split threshold size so only one file need be downloaded to run it. For this file the layer quants are as follows:

embed  : Q3_K
0..7   : alt Q3_K_S Q2_K
8..23  : alt Q3_K_M Q3_K_S
24..44 : Q3_K_M
45     : Q3_K_L
46     : Q4_K_S
47     : Q4_K_M
output : Q5_K

A second smaller quant is also available with predominant Q3_K_S across layers as follows:

(layer distribution updates : v2 5/5/25, v3 5/6/25)
embed  : Q3_K
0..15  : alt Q3_K_S Q2_K
16..38 : Q3_K_S
39..45 : Q3_K_M
46     : Q3_K_L
47     : Q4_K_S
output : Q5_K

This quant reduces file size to 46.6G while maintaining perplexity (10.2) at the same level as a homogenous Q3_K_M quant with ~5G smaller file.

A third quant is also available making heavier use of Q2_K as follows:

embed  : Q3_K
0..15  : alt Q2_K Q2_K_S
16..23 : Q2_K
24..31 : alt Q3_K_S Q2_K
32..43 : Q3_K_S
44..47 : Q3_K_M
output : Q5_K

This quant reduces file size to 42.8 while maintaining good perplexity (10.47) close to homogenous Q3_K_M while being 9G smaller and exhibits no generation artifacts (nonsense words or completing words with chinese characters randomly) across a wide range of test promps.

Comparison:

Quant size PPL Comment
Q2_K 39.6 11.4 many generation artifacts, nonsense words, completes words with chinese tokens randomly
Q2_K_H 42.8e9 10.4 no generation artifacts, no nonsense words, very close to Q3_K_M perplexity
Q3_K_M_3_5 51.6e9 10.2 Q3_K_M with Q5_K output layer
Q3_K_H 46.6e9 10.2 Hybrid quant with Q5_K output layer, Q3_K_M perplexity
Q4_V_H 49.8e9 9.54 Q4_K_H with alt Q2_K on deepest layers for extra early stage nonlinearity
Q4_K_H 50.4e9 9.54 Hybrid quant with Q5_K output layer, IQ4_XS perplexity
Q4_E_H 62.8e9 9.5 E quant with Q6_K embed and output
IQ4_XS 59.9e9 9.57 IQ4_XS with default embed/output layer

Usage Notes:

This model is a vision capable 100B parameter class moe chatbot. It has not been explicitly trained for reasoning with RL (no think block) or optimized for instruction following/agentic tasks/code per state of the art models as of 11/2026. It is multilingual and still may be useful for some stand-alone (no agentic internet access) language oriented tasks which might benefit from its size (~100B class model) while still being runnable on consumer grade hardware. Q4_E_H quant (created 11/2026) is recommended for use, legacy quants generated in the time frame when the model was released are left on the page for now but not recommended for use.

This model can be run on a computer with 64G RAM using combined CPU and a single GPU. The Q2_K_H model might be able to run on a 48G RAM machine but the Q3_K_H and Q4_K_H quants almost certainly won't without either partial offload to a high VRAM GPU or making use of disk swap offload.

A good setup is to offload all model layers to GPU and all non-shared expert FFN tensors to CPU by specifying the override tensor flag: -ot exps=CPU as discussed here : https://github.com/ggml-org/llama.cpp/discussions/13154 This will open up around 45k of KV cache memory on a 12G VRAM GPU at q8_0 KV precision and run around 10t/s on a decent CPU. Note flash attention is needed to run q8_0 KV. To speed up prompt processing under this configuration x2 to x10 you can also either force ubatch (-u) to 16 or use the new --no-op-offload switch added in https://github.com/ggml-org/llama.cpp/pull/13386.

Update: tests on my HW (9900k CPU + 4070 GPU) show that both --no-op-offload and -ubatch 16 are not effective as of llama.cpp b5379.

without --no-op-offload,  ubatch 128
PP=30.856730699382798 TG=9.233060062734445   *best PP and TG config*
with --no-op-offload, ubatch 128
PP=25.74042603622643 TG=8.349090673295033
without --no-op-offload, -ubatch 16
PP=21.448615670596926 TG=9.234865287491097

Therefore it will be necessary to experiment with these flags to determine best settings on different HW configs.

Example conversational prompt (Q4_E_H):

lm how high is up?  Think outside the box.

What a delightfully philosophical and creative question!

When thinking outside the box, "how high is up" becomes a thought-provoking and abstract inquiry that transcends the literal meaning of "up" as a direction in physical space. Here are some unconventional perspectives:

  1. Metaphorical ascension: In a spiritual or emotional context, "up" can represent personal growth, enlightenment, or a higher state of consciousness. In this sense, "how high is up" becomes a question about the limits of human potential, self-improvement, or spiritual evolution. The answer might be "as high as one's aspirations, imagination, or spiritual practice can take them."
  2. Cosmological scales: From an astronomical perspective, "up" can refer to the vast expanse of the universe. The highest "up" might be the edge of the observable universe, which is estimated to be around 93 billion light-years away. However, there may be regions beyond our observable universe, making the concept of "up" even more boundless.
  3. Fractal dimensions: In mathematics, fractals describe self-similar patterns that repeat at different scales. If we consider "up" as a direction in a fractal space, the concept of "how high is up" becomes a matter of scaling. The answer might depend on the specific fractal structure and the level of magnification, leading to an infinite regression of "ups."
  4. Theoretical physics: In theories like string theory or Kaluza-Klein theory, our universe has more than four dimensions. In these frameworks, "up" might refer to movement through additional dimensions, which could be compactified or curled up in ways that defy our classical understanding of space. The height of "up" could be related to the size of these extra dimensions or the energy scales required to access them.
  5. Perceptual relativity: From a subjective perspective, "up" is a relative concept that depends on our frame of reference. For example, in a rotating reference frame, "up" can change direction. Similarly, in a non-Euclidean geometry, like curved spacetime, "up" might not be a fixed direction. This line of thinking leads to the idea that "how high is up" is a matter of perspective, and the answer might vary depending on the observer's context.
  6. Infinity and beyond: Finally, we can consider the notion that "up" is not bound by any physical or mathematical constraints. In this view, "how high is up" is a question that leads to an infinite regress, with "up" becoming a never-ending staircase of possibilities. This perspective invites us to ponder the nature of infinity, the limits of human understanding, and the concept of "beyond" itself.

In conclusion, when thinking outside the box, "how high is up" becomes a rich and multifaceted question that challenges our conventional understanding of space, consciousness, and the nature of reality. The answer, much like the concept of "up" itself, can be seen as relative, subjective, and boundless.

Vision:

Q4_E_H quant goes 1 for 2 on a couple tough bird ID images (F16 vision tower). It does a good job scanning an old IRE journal article by Claude Shannon and summarizing it.

Code:

Q4_E_H quant went 2 for 4 on a small set of code gen prompts; the model does not reliably generate working programs from functional descriptions.

Download the file from below:

Link Type Size/e9 B Notes
Llama-4-Scout-17B-16E-Instruct.Q2_K_H.gguf Q2_K_H 42.8e9 B Good quality
Llama-4-Scout-17B-16E-Instruct.Q3_K_H.gguf Q3_K_H 46.6e9 B Solid quality
Llama-4-Scout-17B-16E-Instruct.Q4_K_H.gguf Q4_K_H 50.4e9 B Excellent quality
Llama-4-Scout-17B-16E-Instruct.Q4_V_H.gguf Q4_V_H 49.8e9 B Extra deep layer nonlinearity for vision mode
Llama-4-Scout-17B-16E-Instruct.Q4_E_H.gguf Q4_E_H 62.8e9 B Runs on 64G Ram machine, minimimum Q4_K across layers
Llama-4-Scout-17B-16E-Instruct.mmproj mmproj 1.75e9 B multimedia projector for vision mode

A discussion thread about the hybrid quant approach can be found here on the llama.cpp git repository:

https://github.com/ggml-org/llama.cpp/discussions/13040

Downloads last month
1,268
GGUF
Model size
108B params
Architecture
llama4
Hardware compatibility
Log In to add your hardware

2-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for steampunque/Llama-4-Scout-17B-16E-Instruct-MP-GGUF