Gemma-4-12B-StyleTune-Minus
Streamlined imatrix quantizations for Gryphe's creative writing architecture. 12B Parameters 2048 Context Window 50 Imatrix CyclesThis repository hosts ik_llama.cpp-compatible GGUF quantizations of Gryphe/Gemma-4-12B-StyleTune. The base model is highly tuned for stylistic flexibility, creative writing, and rich prose. The "Minus" designation on this release refers to the streamlined imatrix generation process—using 50 cycles instead of the heavier 500-cycle runs—to provide a fast, lightweight calibration that still perfectly captures the model's creative essence.
Important: IQK and Trellis quants are ik_llama.cpp-only. They will not load in mainline llama.cpp. If you need a mainline-compatible GGUF, use the BF16 release or convert from the source safetensors.
Quantization Formats
| Quant | Family | Description |
|---|---|---|
| IQ5_K | IQK | Higher fidelity 5-bit. Best quality-to-size for systems with sufficient memory. |
| IQ4_K | IQK | Standard 4-bit IQK. Uses 256-value super-blocks for scaling. |
| IQ4_KS | IQK | "Small" 4-bit variant with tighter packing. Lower memory footprint than IQ4_K. |
| IQ4_KS_R4 | IQK | Row-interleaved variant of IQ4_KS, optimized for ik_llama.cpp's matmul kernels. |
| IQ4_KT | Trellis | Dynamic programming trellis quant. Excellent at minimizing 4-bit error. |
| BF16 | F16 | Full-precision reference. Mainline-compatible. For comparison and conversion. |
Why no MaxOptima? MaxOptima is a highly specific layer-by-layer custom mix. Because StyleTune relies heavily on its unique weight distribution for creative output, a blanket MaxOptima mix hasn't been applied here yet. These standard IQK and Trellis quants offer the most reliable baseline for StyleTune's prose.
Imatrix // The 50-Cycle Trade-off
The importance matrix for these quants was computed from V3 Calibration, derived from OpenIntelligenceNet/Uncensored-Imatrix-Calibration-Mixed-Data. It retains only Uncensored Alpaca and Multi-Turn Roleplay & Fiction, keeping the weights focused strictly on creative and uncensored inference.
| Parameter | Configuration | Impact |
|---|---|---|
| Cycles per chunk | 50 | Significantly faster generation. While 500-cycle runs squeeze out microscopic gains, 50 cycles captures the vast majority of weight importance with a fraction of the compute time. |
| Context window | 2048 tokens | Deep enough to capture multi-turn roleplay dependencies without ballooning RAM usage during quantization. |
| Calibration data | V3 (RP / Uncensored) | Ensures the model's creative writing weights are preserved over general factual knowledge. |
The "Minus" Philosophy: Standard community imatrix runs often use heavy settings (500+ cycles). This release tests a lighter footprint: 50 cycles at 2048 context. The result is a highly efficient calibration process that still targets the exact domains StyleTune excels at. For users who want a lighter, faster download that doesn't sacrifice the model's stylistic soul, this is the optimal path.
Which Quant Should You Choose?
If you are running on limited VRAM or system RAM, start with IQ4_KS or IQ4_KT for the best balance of size and stylistic preservation. If you have ample memory and want to experience the absolute peak of StyleTune's prose capabilities, download IQ5_K. Use BF16 only if you need a mainline reference or plan to convert to another format.
Austriani // Gemma-4-12B-StyleTune-Minus // ik_llama.cpp Base: Gryphe/Gemma-4-12B-StyleTune