@@ -80,6 +80,8 @@ This model is designed as a **coding agent model** for use with Claude Code, Ope
80
- **Use `RYS-Qwen3.5-27B-Uncensored-Splice-IQ4_NL-ik-llama.gguf` if you run ik-llama.cpp.** This is the author's actual daily driver and the variant that will continue receiving the most real-world testing.
81
- **Use `RYS-Qwen3.5-27B-Uncensored-Splice-IQ4_NL-llama.cpp-compatible.gguf` if you run standard llama.cpp or a frontend built on it.**
82
83
## Why IQ4_NL is Recommended Over BF16, Q8, and Q6
84
85
This is not typical. Through extensive real-world testing, the **author's preferred IQ4_NL quantization with custom importance matrix consistently outperformed Q8_0, Q6_K, and full-precision BF16**:
80
- **Use `RYS-Qwen3.5-27B-Uncensored-Splice-IQ4_NL-ik-llama.gguf` if you run ik-llama.cpp.** This is the author's actual daily driver and the variant that will continue receiving the most real-world testing.
81
- **Use `RYS-Qwen3.5-27B-Uncensored-Splice-IQ4_NL-llama.cpp-compatible.gguf` if you run standard llama.cpp or a frontend built on it.**
82
83
+> **Footnote:** The `ik-llama` build is not a uniform IQ4_NL quant. It uses a mixed tensor layout: mostly `iq4_nl`, plus a small number of higher-precision tensors (`iq5_k` / `q6_K`). The `llama.cpp-compatible` build is also mixed, but uses mainline-supported tensor types instead (`q5_K` / `q6_K`).
84
+
85
## Why IQ4_NL is Recommended Over BF16, Q8, and Q6
86
87
This is not typical. Through extensive real-world testing, the **author's preferred IQ4_NL quantization with custom importance matrix consistently outperformed Q8_0, Q6_K, and full-precision BF16**:
