MoQ (Mixture of Quants) is a smart way to shrink AI models without losing their "brainpower." Unlike old methods that treat every part of the model the same, MoQ identifies the most important parts and keeps them high-quality, while heavily compressing the rest to save space.

The result? A model that punches significantly above its weight class.

Benjamin Marie evaluated MoQ GGUFs ("Mixture of Quants") against Unsloth Dynamic (UD) quants, focusing on low-bit versions below 4 bits on average — the range where GGUF models typically struggle most. Results: At similar bits-per-weight (Bpw), MoQ outperforms Unsloth Dynamic quants by ~10% on benchmarks, while also being roughly 2× more token-efficient on average.

"MoQ models are much better than UD quants on benchmarks, and they are also more token-efficient."

Comparison

Here is the comparison between MoQ and Unsloth dynamic quants for LFM 2.5 8BA1B. MoQ perform better i guess .

Thanks to benjamin Marie for his evals

Folder Link	BPW	Total Size
📂 Root	2.75	2.91 GB
📂 Root	3.0	3.17 GB
📂 Root	3.25	3.43 GB
📂 Root	3.5	3.71 GB
📂 Root	3.75	3.96 GB
📂 Root	4.0	4.20 GB
📂 Root	4.25	4.46 GB
📂 Root	4.5	4.72 GB
📂 Root	5.0	5.30 GB
📂 bf16	16.0	16.95 GB

🧠 The MoQ Edge

MoQ optimizes the architecture for the Pareto frontier of memory and performance.

Dynamic Bitrate Allocation: No more "one-size-fits-all." MoQ assigns precision where it actually matters.
Cognitive Preservation: Massive VRAM savings with near-zero degradation in logic and coherence.
Next-Gen Efficiency: Fits "Large" model intelligence into "Small" model hardware.

x : https://x.com/WaleedAhmad1a10 If MoQ does not perform well, email me : waleedahmad.1a10@gmail.com

🛠 Usage & Deployment.

./llama-cli -m Qwen3.5-9B-MoQ-4.0.gguf -p "The future of efficient AI is..."

w-ahmad/LFM2.5-8B-A1B-GGUF-MoQ

🚀 MoQ: Mixture of Quants

Comparison

🧠 The MoQ Edge

🛠 Usage & Deployment.

w-ahmad/
LFM2.5-8B-A1B-GGUF-MoQ