Llama 3.1 Nemotron Nano 8B v1

This compact reasoning model derives from Meta’s Llama 3.1 8B Instruct and is post-trained for conversational work, retrieval-augmented generation, and tool calling. Its target users are teams seeking strong reasoning capability without moving to a much larger model, especially for local deployments and workloads that benefit from efficient inference.

Its post-training combines supervised fine-tuning in mathematics, coding, reasoning, and tool use with reinforcement-learning stages using REINFORCE with RLOO and Online Reward-aware Preference Optimization. The resulting checkpoint aims to improve both reasoning and instruction following, making it a practical fit for local assistants, coding support, retrieval workflows, and other tasks requiring measured reasoning with a long context.