ahmedandaloes/NEXUS-Security-MLX-bf16
ahmedandaloes/NEXUS-Security-MLX-bf16 is a 1.5 billion parameter Qwen2ForCausalLM model, developed by ahmedandaloes, specifically optimized for fast local inference on Apple Silicon. This full-precision (bf16) MLX build retains identical weights to its source, fableforge-ai/NEXUS-Security, and is designed for security-related tasks. It offers a 32768 token context length, making it suitable for detailed analysis in authorized security work.
Loading preview...
NEXUS-Security-MLX-bf16 Overview
This model is an MLX (Apple Silicon) build of the fableforge-ai/NEXUS-Security model, which is based on Qwen2ForCausalLM with approximately 1.5 billion parameters. Developed by ahmedandaloes, its primary purpose is to enable fast local inference on Apple Silicon devices.
Key Characteristics
- Precision: It uses
bf16(brain floating-point) precision, meaning it is a full-precision model with no quantization, retaining 100% of the original weights from the source model. - Size: The model size is approximately 3 GB.
- Context Length: It supports a context length of 32768 tokens.
- Conversion: The model was converted using
mlx-lmandmlxtools, ensuring compatibility and optimization for Apple's Metal Performance Shaders (MPS) framework.
Differentiation and Use Cases
Unlike other quantized versions, this bf16 build offers identical weight fidelity to the original source model, providing maximum accuracy for security-related tasks. While other builds like MLX-4bit, MLX-6bit, and MLX-8bit offer smaller sizes and potentially faster inference at reduced precision, the bf16 version prioritizes full precision for applications where accuracy is paramount. It is specifically intended for authorized security work.
Users should consider this model if they require high-fidelity inference for security applications on Apple Silicon and are willing to utilize the larger file size for full precision.