steampunque/Qwen3.6-35B-A3B-MP-GGUF overview

---

license: apache-2.0

base_model: Qwen/Qwen3.6-35B-A3B

base_model_relation: quantized

tags:

• Qwen3.6 35B A3B
• GGUF
• quantized
• 4 bit
• mixed precision

---

Mixed Precision GGUF layer quantization of Qwen3.6-35B-A3B by Qwen

Original model: https://huggingface.co/Qwen/Qwen3.6-35B-A3B

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. The quants employed are all K to avoid

slow CPU or older GPU processing of IQ quants. An extended layer definition E quant Q4_E_H for the

model is defined as follows (updated 8/26/2026):

   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","VOSD","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_6666"],[1 ,"Q5_K_6665"],[2 ,"Q4_K_6665"],[3 ,"Q4_K_6655"],
   [4 ,"Q4_K_5555"],[5 ,"Q4_K_5555"],[6 ,"Q4_K_5555"],[7 ,"Q4_K_6555"],
   [8 ,"Q4_K_5554"],[9 ,"Q4_K_5554"],[10,"Q4_K_5554"],[11,"Q4_K_6554"],
   [12,"Q4_K_5554"],[13,"Q4_K_5554"],[14,"Q4_K_5554"],[15,"Q4_K_6554"],
   [16,"Q4_K_6554"],[17,"Q4_K_6554"],[18,"Q4_K_6554"],[19,"Q4_K_6554"],
   [20,"Q4_K_6554"],[21,"Q4_K_6554"],[22,"Q4_K_6554"],[23,"Q4_K_6665"],
   [24,"Q4_K_6554"],[25,"Q4_K_6554"],[26,"Q4_K_6554"],[27,"Q4_K_6665"],
   [28,"Q4_K_6554"],[29,"Q4_K_6554"],[30,"Q4_K_6554"],[31,"Q4_K_6665"],
   [32,"Q4_K_6665"],[33,"Q4_K_6665"],[34,"Q4_K_6665"],[35,"Q4_K_6665"],
   [36,"Q5_K_6665"],[37,"Q5_K_6666"],[38,"Q5_K_6668"],[39,"Q6_K_8666"],
   [40,"Q4_K_6554"]
   ]'
   FLAGS="--token-embedding-type Q6_K --output-tensor-type Q6_K --layer-types-high"

The quant was sized to allow running it fully offloaded in 24G VRAM with some room left for both vision tower and context. The layer

quants were optimized for essentially 100% success rate across a curated set of reasoning test prompts using greedy deterministic

sampling and cuda backend using RPC. The minimum quant across layers is Q4_K. The quant includes a ~0.5B MTP layer.

Comparison:

Quant | size | PPL | Comment

-------|--------|-------|-----------

Q4_K_M | 21.7e9 | 6.7 | Q4_K_M with default embedding and output (with ~0.5B MTP)

Q4_E_H | 21.8e9 | 6.7 | Extended K quant with Q6_K embedding and output (with ~0.4B MTP)

Usage:

Qwen3.6-35B-A3B is a vision capable moe RL model. It can be used together with its multimedia projector layers to process images and text inputs

and generate text outputs. The mmproj file is made available in this repository.

Due to the attention scheme used in the model, straightforward speculation approaches cannot be used. Later versions

of llama.cpp as of 4/23/2026 attempt to circumvent this problem however efficiency will not be good so use of speculation

prior to Qwen 3.5 MTP patches is not recommended. As of llama.cpp b9180 MTP support for the model was added to upstream and may be experimented with

by using the Q4_E_H_MTP quant.

The model can be run fully offloaded into 24G VRAM, or with CPU and expert layer offload via config OT="-ot exps=CPU -ngl 99"

Because the model is a 3B active moe the CPU expert offload still gives good gen rate with very large context available.

On a 9900k/4070 or 2x 4070 setup (1 RPC) approx performance (Q4_K_H) is:

CONFIG | QKV | NKV | gen tps

----------------------------|----------|--------|--------

4070+ 9900k CPU exp offload | F16 | 470k+ | 31

4070+ 9900k CPU exp offload | Q8_0 | 832k | 31

2x4070 (RPC) | F16 | 32k + | 85

2x4070 (RPC) + vision tower | F16 | 32k + | 88-91

2x4070 (RPC) | Q8_0 | 64k + | 66

High context yarn config is as follows: set base context for yarn rope scale compute to 262144 (256k), then with a context of 470k tokens the rope scale = 470 / 256 = 1.83.

Then on model start pass --rope-scaling yarn --yarn-orig-ctx 262144 --rope_scale 1.83 (must be ajusted if kv other than 470k)

Later versions of llama.cpp have a bug which soft caps context length to the training context, effectively disabling yarn context extension. Patch server-context.cpp according to https://github.com/ggml-org/llama.cpp/issues/22140 to fix it.

The model appears to be trained to decide itself whether to do a think block or not. When it does a think block it can fall into very

heavy overthinking on some prompts. Over a curated set of eval prompts the model did exceptionally well, missing only one IQ test prompt

in both think and nothink modes. To avoid overthinking inject think start and think stop tokens first thing after assistant prompt:

THINK_START="<think>\n"
THINK_STOP="\n</think>\n\n"

If the model doesnt feel like doing thinking on a given prompt it will automatically do this. To force the model into

a think block inject a bootstrap think start following the assistant prompt:

"<think>\n"

VISION:

The quant was tested in vision mode on a couple pretty tough bird ID image and did well, with accurate and concise answers on both tough prompts.

CODE:

The quant was tested across a small set of code gen prompts and found to be capable of generating working code

with or without think block enabled on some prompts, but failed on others. Overall it was found not to be robust at code gen

across the tiny set of test prompts given.

This version of the model needs a verion of llama.cpp which supports MTP due to included MTP layer; if not using MTP warning messages will be generated about unused tensors but the model will run.

Benchmarks:

A full set of both math and vision benchmarks for the model will eventually be given here: https://huggingface.co/spaces/steampunque/benchlm

Download the file from below:

| Link | Type | Size/e9 B | Notes |

|------|------|-----------|-------|

| Qwen3.6-35B-A3B.Q4_E_H.gguf | Q4_E_H | 21.8e9 B | ~Q4_K_M size, includes ~0.4B MTP layer |

| Qwen3.6-35B-A3B.mmproj.gguf | F16 | 0.90e9 B | multimedia projector |

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

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