A 27B parameter model reasoning-distilled from Claude 4.6 Opus, quantized to IQ4_XS with importance matrix for optimal quality/size tradeoff. The v2 training improves tool calling accuracy by 31.6% over v1 on quantized models.

Files

File	Size	Description
`Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2.IQ4_XS.gguf`	14.7GB	IQ4_XS imatrix quantization

Performance

Tested on dual AMD Radeon RX 7900 XTX (2× 24GB VRAM):

~30 tok/sec generation
131K context window
Tool calling confirmed working

Usage

llama.cpp

```bash llama-server
-m Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2.IQ4_XS.gguf
--host 0.0.0.0 --port 8080
-c 131072 -ngl 99
--think ```

vLLM

Not recommended — use the FP8 variant for vLLM.

Quantization Details

Source: v2 BF16 weights
Method: IQ4_XS with importance matrix (imatrix)
Imatrix: mradermacher/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-i1-GGUF
Tool: llama.cpp

Credits

Original model: Jackrong
Imatrix calibration: mradermacher / nicoboss
Quantization: SolidRusT Networks

solidrust/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF

Qwen3.5-27B Claude 4.6 Opus Reasoning Distilled v2 — GGUF

What This Is