Llama 3.3 Nemotron Super 49B v1.5

Llama 3.3 Nemotron Super 49B v1.5 is NVIDIA's post-trained derivative of Meta's Llama 3.3 70B, designed for reasoning-intensive tasks while staying memory-efficient. Using Neural Architecture Search with block-wise distillation, it reduces the footprint of the base 70B model while preserving quality, enabling deployment on a single high-end GPU. The model performs strongly on mathematical reasoning (97.4% on MATH500, 87.5% on AIME 2024) and code generation (73.58% on LiveCodeBench), and supports tool-calling for agent systems. It offers 128K token context and operates in two modes, reasoning-enabled with chain-of-thought and reasoning-disabled for faster inference. Trained on math, code, science, and domain-specific data via supervised fine-tuning and reinforcement learning, it handles instruction-following, retrieval-augmented generation, and complex problem-solving. Reasoning mode is controlled via system prompt, with a recommended temperature of 0.6 and top-p of 0.95.