NVIDIA Nemotron 3 Ultra showcasing advanced AI models NeMo Automodel, Megatron Bridge, and reinforcement learning recipes for

Editorial illustration for NVIDIA Nemotron 3 Ultra adds NeMo Automodel, Megatron Bridge and RL recipes

NVIDIA Nemotron 3 Ultra adds NeMo Automodel, Megatron...

NVIDIA Nemotron 3 Ultra adds NeMo Automodel, Megatron Bridge and RL recipes

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

June 4, 2026 • 2 min read

Single‑turn chatbots are giving way to long‑running agents that can reason, keep context, call tools and hand off work to sub‑agents. While that flexibility expands what AI can do, each interaction adds tokens, and token counts climb quickly as agents plan, invoke tools, receive data and feed history back into the model. The result? Higher costs and a greater chance that the system veers off its original goal.

Developers are responding with a two‑tiered approach: frontier‑reasoning models handle orchestration and complex planning, while leaner models take care of high‑volume execution, validation and tool calls. NVIDIA’s latest offering, Nemotron 3 Ultra, slots into the first tier. It is a 550‑billion‑parameter Mixture‑of‑Experts model, but only 55 billion parameters are active at any moment.

NVIDIA says the model can sustain “hard calls” such as architectural decisions across coding sessions, synthesizing contradictory evidence from hundreds of research sources, or verifying chip designs against thousands of constraints. In benchmark tests on SWE‑bench and Terminal bench 2.0, Nemotron 3 Ultra reportedly processes five times as many tokens per second as comparable open models and does so with fewer total tokens and fewer tokens per turn.

Nemotron 3 Ultra adds Multi-Teacher On-Policy Distillation Multi-Teacher On-Policy Distillation (MOPD) is a training method in which Ultra learns from multiple specialized teacher models while generating its own attempts during training.

— Ultra, NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents - NVIDIA Developer Blog

Why this matters We see NVIDIA extending Nemotron 3 Ultra with a suite of NeMo recipes that target the growing demand for long‑running agents. The added Automodel LoRA and Megatron Bridge aim to cut inference latency while keeping context over many turns, a claim that could ease the token‑bloat problem developers face today. Yet the article notes that continuous tool calls and sub‑agent invocations still swell token counts, driving up costs and raising the specter of goal drift.

Our community must ask: can these new RL GRPO and MOPD recipes really curb that drift, or simply shift the burden to more complex fine‑tuning pipelines? The Dynamo deployment recipe promises smoother rollout, but the walkthrough stops short of showing real‑world scaling results. For founders eyeing commercial agents, the promise of faster reasoning is appealing; for researchers, the breadth of recipes offers a sandbox for experimentation.

Unclear whether these additions will translate into measurable savings in production environments, but they give us a concrete set of tools to test the hypothesis that longer‑running agents can stay efficient without sacrificing reliability.