NVIDIA has released Llama‑3.3‑Nemotron‑Super 49B v1.5, an open-weight LLM designed to deliver top-tier reasoning, math, and tool-calling performance at a mid-size model scale. The model outperforms leading open competitors like Qwen3‑235B, DeepSeek R1‑671B, and even NVIDIA’s own prior Nemotron Ultra 253B across key reasoning benchmarks. Despite its smaller footprint, it features a 128K token context window and excels at multi-turn reasoning tasks.

What sets v1.5 apart is its combination of size and accessibility. It was built using neural architecture search (NAS) to optimize for H100/H200 GPUs, meaning it runs efficiently on a single high-end card. This lowers the barrier for developers building RAG agents, math solvers, and code assistants in real-world applications. Alongside the model, NVIDIA has released the full post-training dataset used for alignment and reasoning tuning, a move that enhances transparency and reproducibility in commercial deployments.

In a field increasingly dominated by massive, inaccessible models, NVIDIA is positioning Nemotron Super v1.5 as the pragmatic choice for agentic system developers. It's not just a benchmark leader. It's designed to work in actual production environments, with open weights, permissive licensing, and GPU efficiency that SMBs and startups can use today.