Nvidia launches Nemotron 3 open reasoning models with Nano (30B), Super (100B), and Ultra (500B) variants. 4x higher throughput and 1M token context window.
What Nemotron 3 Brings to Agentic Workloads
Nvidia’s Nemotron 3 family targets a different problem than general chat: long-running agent loops that plan, call tools, check intermediate results, and revise. Those loops need strong multi-step reasoning, room for large tool traces, and enough speed that an agent can iterate without stalling a product. Nemotron 3 is positioned as open reasoning models built for that pattern—models you can run, inspect, and wire into pipelines rather than treat as a sealed API.
Open weights matter here because agent systems are rarely one-shot prompts. Teams tune system instructions, constrain tool schemas, add retrieval, and evaluate failure modes on their own data. An open reasoning stack lets you control latency budgets, data residency, and fallback paths when a tool call fails or a plan goes off track.
Three Scale Points: Nano, Super, and Ultra
The lineup is organized by size so you can match model capacity to task difficulty and hardware cost:
- Nano (30B) — a smaller option for high-volume agents, edge or single-GPU serving, and tasks where most steps are routine (routing, short plans, light tool use).
- Super (100B) — a mid-tier choice when plans get longer, tools return noisy data, or the agent must hold more intermediate state without collapsing into shallow answers.
- Ultra (500B) — the largest variant for hard reasoning, multi-hop research, and workflows where quality per turn outweighs raw request volume.
In practice, treat the three sizes as a ladder, not a single default. Many production systems run Nano for classification and tool selection, escalate ambiguous or high-stakes steps to Super or Ultra, and keep the expensive path only where evaluation shows a clear gain.
Throughput and a 1M-Token Context Window
Nemotron 3 is described with roughly 4x higher throughput relative to prior expectations for this class of work, and a 1M token context window. For agents, those two properties interact. Higher throughput shortens the time between “think → act → observe,” which is where user-facing latency and cost often hide. A million-token window is less about stuffing more prose for its own sake and more about keeping full trajectories—prior plans, tool JSON, logs, and retrieved docs—inside one coherent context instead of aggressive truncation that erases the evidence the model needs to recover from mistakes.
Still, a large window is not free. Long contexts raise memory pressure, can dilute attention if filled with noise, and make debugging harder if you never prune. Use the 1M capacity for structured state: summarized history, the current plan, the last few tool results, and only the retrieval slices that justify the next action. Measure tokens per successful task, not tokens per request.
How to Evaluate and Ship an Agent on Nemotron 3
Start with a fixed task suite that mirrors real failure modes: multi-tool sequences, partial tool failures, long-horizon goals, and cases where the model should refuse or ask for clarification. Score plan validity, tool-call correctness, and final outcome—not just fluent prose. Run the same suite across Nano, Super, and Ultra so you see where size stops buying accuracy.
On the serving side, design for escalation and observability. Log each reasoning segment and tool exchange with stable IDs, cap concurrent long-context jobs so throughput stays predictable, and keep a smaller model in the hot path for cheap steps. Because the models are open, you can pin a revision, A/B test prompts and tools offline, and redeploy without waiting on a vendor’s closed schedule. The win is not novelty—it is a clearer fit between model size, context, and the iterative nature of agentic AI.