A massive mixture-of-experts model with 397 billion total parameters but only 17 billion active at a time, meaning it punches above its computational weight by routing each token through a selective subset of its capacity. It runs in GGUF format, making it accessible for local deployment despite its scale. The quantization to Q4_K_XL keeps memory footprint manageable while preserving much of the full-precision model's behavior.