Update README.md

README.md

@@ -17,6 +17,15 @@ tags:
 These models are exl3 quantization models of [Qwen2.5-Coder-32B](https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct) which is still SOTA no-reasoning coder model as of today. This model is still my go-to FIM(fill in the middle) autocompletion model after Qwen3, Gemma3 release.
 I used [exllamav3 version 0.0.2](https://github.com/turboderp-org/exllamav3/releases/tag/v0.0.2).
 
+## EXL3 Quantized Models
+
+[4.0bpw](https://huggingface.co/LLMJapan/Qwen2.5-Coder-32B-Instruct_exl3/tree/4.0bpw)
+[6.0bpw](https://huggingface.co/LLMJapan/Qwen2.5-Coder-32B-Instruct_exl3/tree/6.0bpw)
+[8.0bpw](https://huggingface.co/LLMJapan/Qwen2.5-Coder-32B-Instruct_exl3/tree/8.0bpw)
+
+For coding, I found >6.0bpw or preferably 8.0bpw model with KV Cache Quantization (>6.0bpw) is much better than 4.0bpw.
+If you are using these models only for short Auto Completion, 4.0bpw is usable.
+
 ## Credits
 
 Thanks to excellent work of exllamav3 dev teams.