islameissa/Qwen3.6-27B_IQ4_GGUF

Qwen 3.6-27B Custom Imatrix Quants (IE-Series)

This repository provides highly optimized GGUF quantizations of Qwen 3.6-27B using advanced imatrix (Importance Matrix) calibration.

While the initial Q4_0-IE provided a solid baseline, the newly developed Q4H variant represents a significant jump in quality, outperforming standard Q4_K_M and Q4_1 quants in both perplexity and stability (KLD P99).

Methodology: The ActivEnergy "IE" Calibration

These quants were generated using a custom-curated calibration dataset. I began with the Bartowski Calibration v5 text as a foundation and integrated modifications designed to better capture high-precision logic and nuanced language patterns.

• Tool: llama.cpp / llama-quantize
• Calibration: Hybrid IE-Custom + some of Bartowski v5 calibration text
• Primary Objective: Minimize KLD (Kullback–Leibler Divergence) outliers, particularly in the P99 range, to ensure model stability during long-form generation.

Performance & Compatibility

• Speed: Tested on RTX 5090, achieving approximately 60 TPS.
• Agents: Fully compatible with multi-agent frameworks like OpenClaw.

---

Evaluation Report: ActivEnergy Quantizer

The following metrics were derived using the Qwen 3.6-27B architecture as the base. The IE-Q4H variant is the current top recommendation for users seeking the best balance between size and logical fidelity.

Comparison Table

Rank	Variant	File Size	Mean PPL	KLD Mean	KLD Median	KLD P99
--	F16 Baseline	50.1 GB	2.0878	N/A	N/A	N/A
1	IE-Qwen36-27b-Q4H	16.9 GB	2.0844	0.010602	0.002350	0.084462
2	Unsloth-UD-Q4_K_XL	16.4 GB	2.0716	0.008079	0.002192	0.146924
3	IE-IQ4_NL	14.7 GB	2.1352	0.019256	0.002533	0.175252
4	IE-IQ4_XS	14.0 GB	2.1263	0.018374	0.002328	0.184353
5	Bartowski-Q4_1	16.4 GB	2.0844	0.012945	0.002636	0.219050

Happy Inference! Please leave a comment with your results or any testing data.