Quantization Shootout Results
I ran a quantization shootout on Qwen3-Coder-Next to test different precision levels. The goal was to see how these variations affect both quality and performance.
Hardware and Setup
- Hardware: 3× R9700 PRO (96 GB VRAM)
- Backend: llama.cpp Vulkan
- Evaluation: wikitext-2 (583 chunks, ctx 512 tokens)
- Formats Tested: MXFP4_MOE Q4_K_M Q5_K_M UD-Q5_K_M
The Results
The UD-Q5_K_M format stood out as having the best balance between quality and file size. It achieved a 94% top-1 accuracy, which is better than both Q4_K_M and Q5_K_M by significant margins.
| Metric | MXFP4 | Q4_K_M | Q5_K_M | UD-Q5_K_M |
|---|---|---|---|---|
| Same top-1 | 89.4% | 89.6% | 93.0% | 94.0% |
| Mean KL divergence | 0.0746 | 0.0685 | 0.0308 | 0.0217 |
| Max KL (worst token) | 13.04 | 5.93 | 8.19 | 4.75 |
| File size | 44.7 GB | 45.2 GB | 52.9 GB | 55.2 GB |
UD-Q5_K_M wins on every quality metric tested.
A 5% difference in per-token agreement becomes a 500× difference by token 100, highlighting the importance of token accuracy in long sequences. MXFP4 had a 94.0% top-1 accuracy compared to UD-Q5_K_M’s 94%, but it was still slightly ahead in this specific test.
The file size difference between MXFP4 (44.7 GB) and UD-Q5_K_M (55.2 GB) is notable, with UD-Q5_K_M being ~10 GB larger than MXFP4.
Performance
- Refill (batch 512): MXFP4 was still the fastest due to hardware kernels.
- Prefill (batch 4096): MXFP4 maintained its lead.
- Decode: Q4_K_M edges UD-Q5 slightly, but UD-Q5 is within 9% despite being 22% larger.
For interactive coding tasks (which are decode-bound), the speed hit from using UD-Q5_K_M was negligible. I have since switched my default format to UD-Q5_K_M for daily code generation, as it offers a better quality-to-size trade-off compared to MXFP4.
Quants for Code Models
What quantization techniques are you using for your code models? Have you noticed similar quality cliffs with aggressive compression?
Key Takeaways
- The UD-Q5_K_M format provides the best balance between quality and file size.
- A 5% difference in token accuracy can lead to a significant change, especially in long sequences.
- For interactive tasks like code generation, the speed hit from using lower precision models is often minimal.
