New Benchmarks for CyberTiel 35B-A3B Model
An anonymous researcher has released performance benchmarks for CyberTiel 35B-A3B, a quantized large language model designed for software engineering and offensive security research. According to reports circulating on Reddit, the model is a specialized version of the TielCoder base model, utilizing a 4-bit quantization method with an improved “imatrix” derived from a curated corpus of cybersecurity and agentic software engineering data.
The developer claims that the model’s “abliteration” process—intended to remove censorship—has a negligible impact on performance under Q4 quantization. The researcher asserts that the model preserves the high-precision weights necessary for complex coding tasks while allowing for legitimate security research that censored models might restrict.
Performance Metrics and Benchmarking
Unverified benchmark data shared by the developer suggests that CyberTiel 35B-A3B outperforms several existing models in specific coding environments. In tests involving real codebase issues, the model reportedly outperformed the Opus 4.6 medium model. Furthermore, the developer claims the model operates at a significantly higher speed than the Qwen3.8-27b medium model, allegedly resolving three issues in the time it takes the Qwen model to solve one.
The model is specifically optimized for the SWE-bench-Live testing framework, which evaluates a model’s ability to solve real-world software engineering problems without breaking existing code. However, the developer notes a trade-off in the model’s architecture: the 35B-A3B parameter size sacrifices general world knowledge to prioritize specialized coding capabilities.
Technical Specifications and Availability
The model is currently available in GGUF and MLX formats via Hugging Face. Users have noted that the model’s performance is highly dependent on hardware configurations and quantization settings. In community discussions, users have discussed various implementation methods, including the use of speculative decoding (draft-dflash) and Multi-Token Prediction (MTP) to optimize inference speeds.
While the developer emphasizes that the model provides increased user control and alignment by removing third-party censorship, some community members have raised questions regarding the practicalities of running such models on consumer-grade hardware, such as the NVIDIA 3060Ti.
Industry Context and Safety Discussions
The release of specialized, uncensored models like CyberTiel occurs amid broader industry debates regarding AI alignment and the risks of autonomous systems. While some observers argue that censorship impedes legitimate professional work, others have raised concerns regarding the long-term implications of AI becoming deeply embedded in critical infrastructure, such as power grids and financial systems.
As specialized models continue to emerge in the open-source community, the distinction between general-purpose AI and task-specific tools—such as those used in CAD or circuit design—remains a point of technical and philosophical discussion within the field.