Standard quantization rounds all weights equally. iMatrix runs a calibration pass over real text to identify which weights matter most, then prioritizes precision where it counts. The result: noticeably better coherence and instruction-following at Q2/Q3/Q4 — same file size, better output.

The i-quants (IQ2_M, IQ3_M, IQ4_XS) are exclusively iMatrix-based and provide the best quality-per-GB available.

Quick Start

Ollama

ollama run hf.co/liodon-ai/Qwen3.6-35B-A3B-imatrix-GGUF:Q4_K_M

llama.cpp

llama-cli -hf liodon-ai/Qwen3.6-35B-A3B-imatrix-GGUF:Q4_K_M

LM Studio / Jan

Search liodon-ai/Qwen3.6-35B-A3B-imatrix-GGUF and pick your quant.

Available Quants

Quant	Size	Notes
`Q2_K`	12.94 GB	tiniest standard — runs almost anywhere
`Q3_K_M`	16.76 GB	great for 8GB VRAM
`Q4_K_M`	21.17 GB	sweet spot (recommended)
`Q5_K_M`	24.73 GB	high quality
`Q6_K`	28.51 GB	near-lossless
`Q8_0`	36.90 GB	basically full quality

VRAM Requirements

VRAM	Recommended Quant
6 GB	`IQ2_M`
8 GB	`IQ3_M` or `Q3_K_M`
10 GB	`IQ4_XS` or `Q4_K_M`
12 GB	`Q4_K_M`
16 GB	`Q5_K_M`
24 GB	`Q6_K` or `Q8_0`

iMatrix vs Standard — Why It Matters

At low bit widths (Q2/Q3/Q4), standard quantization loses coherence and starts producing repetitive or broken output. iMatrix keeps the model sharp by protecting the most important weights. If you're running at Q4 or below, prefer the iMatrix quants from this repo over standard Q-series from other repos.

Base Model

Model: Qwen/Qwen3.6-35B-A3B
Params: 35B (3.5B active, MoE)
Context: 128K tokens
License: Apache 2.0
Authors: Qwen Team (Alibaba Cloud)