MoQ: Mixture of Quants
🚀 MoQ: Mixture of Quants
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.**Stop settling for uniform bitrates. Standard quantization is a relic of the past, treating vital cognitive weights the same as redundant noise. **
The result? A model that punches significantly above its weight class.
Comparison
Here is the comparison between MoQ and Jackrong's quants for his model. MoQ perform better by such a big margin that you can save a GB for same performace . All 3 metrics prove how better they are :
Background evaluations:
Benjamin Marie evaluated MoQ GGUFs ("Mixture of Quants") against Unsloth Dynamic (UD) quants on original Qwen 3.5 9B, 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."
Files
| Folder Link | BPW | Total Size | Description |
|---|---|---|---|
| 📂 MoQ-Quants | 3.3 | 3.83 GB | |
| 📂 MoQ-Quants | 3.7 | 4.28 GB | |
| 📂 MoQ-Quants | 3.9 | 4.47 GB | |
| 📂 MoQ-Quants | 4.2 | 4.89 GB | |
| 📂 MoQ-Quants | 4.4 | 5.09 GB | |
| 📂 MoQ-Quants | 4.7 | 5.36 GB | |
| 📂 MoQ-Quants | 4.9 | 5.62 GB | |
| 📂 MoQ-Quants | 5.0 | 5.74 GB | |
| 📂 MoQ-Quants | 5.2 | 6.00 GB | |
| 📂 MoQ-Quants | 5.4 | 6.17 GB | |
| 📂 MoQ-Quants | 6.6 | 7.62 GB |
This is the MTP Repo . MTP speculative decoding is for faster generation
🧠 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-4.85.gguf -p "The future of efficient AI is..."