NVIDIA Nemotron 3

What It Is

NVIDIA Nemotron 3 is an open-weight model family purpose-built for agentic AI. The architecture uses mixture-of-experts with 120 billion total parameters but only 12 billion active per inference, making it efficient for deployment. The family targets three domains: robotics (with Agility Robotics, Figure AI, Skild AI as early adopters), enterprise workflows (ServiceNow, Uber), and developer tooling (Cursor, Perplexity). The models achieved 85.6% on PinchBench and are expected to reach general availability in the first half of 2026.

Why It Matters

Nemotron 3 enters at Watch because it's pre-GA, but NVIDIA's entry into the open model space deserves attention. The thesis is different from Meta's Llama or Google's Gemma: Nemotron is designed from the ground up for agentic workloads — tool use, planning, and multi-step execution. The mixture-of-experts architecture (12B active out of 120B total) is explicitly an efficiency play for deployment at the edge and in enterprise environments.

The early adopter list is the interesting signal. Cursor and Perplexity choosing Nemotron alongside their existing model options suggests it offers something useful for agentic coding and research workflows. NVIDIA's GPU monopoly means they can optimise the models for their own hardware in ways other model providers can't match.

Key Developments

• Mar 2026: NVIDIA announces Nemotron 3 family at GTC 2026. Open weights, 120B params / 12B active (MoE).
• Feb 2026: Early adopter announcements: Cursor, Perplexity, ServiceNow, Uber, Agility Robotics, Figure AI.
• Feb 2026: 85.6% on PinchBench. Expected GA first half 2026.

What to Watch

GA release timing and benchmark performance against established open models (Llama 4, DeepSeek) will determine whether this is a meaningful player or a GPU-company side project. Watch for NVIDIA-specific optimisations — if Nemotron runs significantly faster on NVIDIA hardware than competing open models, it creates a hardware-model bundle that's compelling for enterprises already committed to NVIDIA GPUs. The robotics angle (Agility, Figure AI) is a differentiated use case worth tracking separately.