The "Heretic" series is an abliterated/uncensored version of Google's Gemma 4 26B MoE model, specifically tuned to bypass refusal vectors while maintaining high reasoning and coding capabilities.

Model Highlights

Architecture: Mixture-of-Experts (MoE) - 26B total parameters, ~3.8B active per token.
Uncensored: Significantly reduced refusal rate (from 100/100 to 11/100 in internal testing).
Intelligence Retention: Maintains a 91.73% PIQA accuracy, nearly identical to the base model's 92.06%.
Speed: Capable of 90+ t/s on consumer hardware (tested on RTX 5060 Ti 16GB).

Performance Stats (Abliteration)

Metric	Heretic	Original Model
KL Divergence	0.0468	0 (by definition)
Refusals	✅ 11/100	❌ 100/100
PIQA Accuracy	91.73%	92.06%

Quantization Details

These quants were created using llama.cpp and optimized for various VRAM configurations. IQ4_NL is highly recommended for 16GB VRAM cards (like the 5060 Ti), as it provides the best balance of speed and "Non-Linear" weight preservation.

IQ4_NL: 14.6 GB - Recommended for 16GB VRAM.
Q4_K_M: 16.8 GB - Standard balanced quant.
Q6_K: 22.6 GB - Near-lossless quality for 24GB+ VRAM.
Q3_K_L: 13.8 GB - Maximum VRAM savings for 12GB-16GB cards.

🚀 Quick Start: Optimized for 16GB VRAM (RTX 5060 Ti) To achieve 90+ t/s and a 64k context window on a single 16GB card, use the following llama-server configuration. This setup leverages the Gemma 4 TurboQuant update, allowing for near-lossless performance even with quantized KV caches.

PowerShell

.\llama-server.exe `
  -m .\gemma-4-26b-heretic.IQ4_NL.gguf `
  --host 0.0.0.0 `
  --port 8080 `
  -c 65536 `
  --n-gpu-layers 99 `
  --flash-attn on `
  --cache-type-k q4_0 `
  --cache-type-v q4_0 `
  --reasoning-budget 1024

Why this config works: KV Cache Q4_0: Thanks to Gemma 4's architecture and recent llama.cpp optimizations, using 4-bit KV cache is virtually lossless. This allows the 64,000 token context to fit comfortably alongside the model weights. Reasoning Budget: While the model can handle 4096, a 1024-2048 budget is recommended for standard users to maintain blistering generation speeds while still allowing for deep logical "thinking" on complex prompts. Full GPU Offload: --n-gpu-layers 99 ensures the entire 26B-A4B MoE structure (only ~3.8B active params) stays on the VRAM bus for maximum throughput.

Acknowledgments

Original Model: llmfan46
Abliteration Method: Orthogonalization / Targeted attn.o_proj intervention.
Quantized by: Vastopian

Vastopian/gemma-4-26B-A4B-it-heretic-GGUF

Gemma-4-26B-A4B-it-Heretic GGUF

Model Highlights

Performance Stats (Abliteration)

Quantization Details

Acknowledgments

Vastopian/
gemma-4-26B-A4B-it-heretic-GGUF