CyberTiel 35B-A3B Uncensored 4-Bit Model Delivers Fast, High-Performance Local Coding
An independent researcher released CyberTiel 35B-A3B, an uncensored 4-bit quantized local language model optimized for agentic software engineering and cybersecurity tasks. Using a custom importance matrix (imatrix), the model outperforms Opus 4.6 medium on real-world SWE-bench-Live codebase issues while running in a fraction of the time required by Qwen3.8-27b medium. This release demonstrates that carefully applied quantization with custom imatrix calibration can offset performance degradation caused by uncensoring (abliteration). It empowers open-source developers and security researchers with a fast, locally hostable model capable of handling offensive security tasks and real-world coding without vendor censorship restrictions. The model was quantized to 4-bit (Q4) using an imatrix trained on a curated corpus of cybersecurity and agentic engineering workflows to preserve vital weights. Available on Hugging Face in GGUF and MLX formats, CyberTiel sacrifices some general world knowledge to maximize pure coding speed and issue-resolving capabilities.
## BACKGROUND
Uncensoring models via abliteration modifies internal weights to eliminate refusal behaviors, which can degrade reasoning performance if done improperly. Quantization reduces LLM memory requirements by converting floating-point weights to lower precision formats like 4-bit, while an importance matrix (imatrix) measures activation importance to minimize quality loss during this compression process.