---
base_model: Qwen/Qwen3.5-122B-A10B
tags:
- gguf
- quantization
- iq4_ks
- iq4_k
- ik_llama.cpp
- qwen
- qwen3_5_moe
pipeline_tag: text-generation
---

# Qwen3.5 122B A10B - Custom GGUF Quantizations

🚨 **CRITICAL COMPATIBILITY WARNING** 🚨
**These are `iqk` format quantizations and are EXCLUSIVE to the `ik_llama.cpp` fork.** They will **NOT** work on mainline `llama.cpp`, standard LM Studio, standard Text Generation WebUI, or KoboldCPP. You *must* compile and run this using [ikawrakow's llama.cpp fork](https://github.com/ikawrakow/llama.cpp) (or a UI where you have manually swapped the backend to an `ik_llama` build).

---

This repository contains custom, mixed-precision `ik_llama.cpp` GGUF quantizations for [Qwen/Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B). 

These quants were specifically designed to push the routed expert layers to slightly higher precision (`IQ4_KS` and `IQ4_K`) while maintaining strict precision on the attention and embedding layers.

## ⚠️ Disclaimer: The "Vibes Test"
**These quantizations have NOT been formally tested for perplexity.** They were compiled blindly as an experiment to see how the model handles shifting bottlenecks. There is no guarantee that they are mathematically optimal or perform flawlessly. They are provided entirely as-is. If they pass the vibes test for you, enjoy! 

## 🙏 Credits & Acknowledgments
Massive credit goes to [ubergarm/Qwen3.5-122B-A10B-GGUF](https://huggingface.co/ubergarm/Qwen3.5-122B-A10B-GGUF). 
The `imatrix.dat` used to calculate these custom quants was pulled directly from their phenomenal enterprise-hardware run, and the custom layer-mapping recipes used here are heavily based on their original blending logic.

---

## 🛠️ Quantization Recipes

### 1. The IQ4_KS Mix
This mix balances an upgraded routed-expert layer with highly compressed (but imatrix-optimized) embeddings to save VRAM.
* **Token Embeddings & Output:** `IQ6_K`
* **Attention / Delta Net / Shared Experts:** `Q8_0`
* **Routed Experts:** `IQ4_KS`

### 2. The IQ4_K Mix
This mix opts to spend a tiny bit more VRAM to give the model absolute `Q8_0` precision on its vocabulary, alongside slightly heavier experts.
* **Token Embeddings & Output:** `Q8_0`
* **Attention / Delta Net / Shared Experts:** `Q8_0`
* **Routed Experts:** `IQ4_K`

---

## 💻 How to Run

1. Clone and build the `ik_llama.cpp` fork from [ikawrakow/ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp).
2. Use the compiled `llama-server` or `llama-cli` from that specific build.

**Example `llama-server` launch command:**
```bash
./llama-server -m Qwen3.5-122B-A10B-IQ4_KS.gguf -c 8192 -ngl 99 -fa