NVIDIA NeMoTron 3 Ultra: Blazing Fast, Great at Agentic Tasks, Inconsistent at Code

NVIDIA dropped a bombshell on June 4, 2026, with NeMoTron 3 Ultra — a 550-billion-parameter open-weight model redefining what’s possible in agentic AI. But this isn’t a typical dense giant. Ultra is a carefully engineered beast whose strengths and weaknesses reveal where large language models are heading.

What Makes Ultra Different from Super?

The NeMoTron 3 family spans three tiers — Nano (30B), Super (120B), and Ultra (550B). But the jump from Super to Ultra isn’t just about scale. Super is a 120B dense model — all parameters active on every forward pass. Ultra uses a LatentMixture-of-Experts (LatentMoE) architecture with 550B total parameters but only 55B active per token. That’s 9.2× more total parameters than Super, but only 4.6× more compute per inference.

The real architectural innovation is the Mamba-2 + Attention hybrid: interleaved state-space model (Mamba-2) layers, MoE layers, and selective attention layers work together to balance speed and quality. Multi-Token Prediction (MTP) enables native speculative decoding with five speculative tokens, dramatically accelerating inference throughput. The model was pre-trained natively in NVFP4 (4-bit floating point) — not quantized after the fact, but trained from scratch at this precision — giving it a significant efficiency edge over post-training quantization approaches.

Benchmark Dominance — Especially in Agentic Tasks

The numbers tell a clear story. On SWE-Bench Verified, Ultra scores 70.7 (BF16) vs. Super’s 60.47. But the standout is Terminal Bench 2.1: Ultra hits 56.4, nearly double Super’s 31.0 — a paradigm shift for autonomous agent tasks.

Ultra also shines across agentic benchmarks: TauBench V3 averages 70.3 (with standout scores of 93.6 in telecom and 88.4 in retail), PinchBench reaches 89.8, and the model handles up to 1 million tokens of context — essential for long agent trajectories and complex multi-turn interactions.

The configurable reasoning mode (enable/disable via chat template) allows users to trade speed for depth depending on the task, a flexibility that NVIDIA’s technical report highlights as a key design principle.

Real-World Performance: Fast, but Inconsistent at Code

In practice, NeMoTron 3 Ultra delivers on its speed promises. The Mamba-2 hybrid architecture, LatentMoE sparsity, and MTP speculative decoding make it one of the fastest models in its weight class.

Where Ultra truly excels is agentic system tasks: fixing broken package installations, organizing directories, orchestrating DevOps workflows, and navigating shell environments autonomously. The 2× gain on Terminal Bench translates directly to real productivity.

However, pure code generation tells a different story. While Ultra excels on SWE-Bench — editing and iterating on existing codebases with agentic loops — it struggles with generating clean, correct code from scratch. Independent testing shows the model produces inconsistent results on standalone coding benchmarks, occasionally missing edge cases or producing logically flawed implementations. The blazing speed sometimes comes at the cost of attention to detail in nuanced programming tasks.

This creates an interesting profile: Ultra is better as an AI coding agent that iterates on existing code than as a standalone code generator.

Best Use Cases

NeMoTron 3 Ultra is purpose-built for agentic AI:

• Autonomous DevOps and system administration — shell troubleshooting, installation repair, infrastructure management.
• File organization and data pipeline orchestration — multi-step operations across large directory structures.
• Complex multi-turn agent workflows — tasks requiring search, tool calling, and context gathering over long trajectories.
• Long-context analysis — the 1M token window handles entire codebases in a single pass.
• Software engineering agents — IDE-integrated assistants that edit, refactor, and debug existing codebases.

For pure code generation from scratch or one-off scripting, users may find more consistent results with smaller, code-specialized models. But for agentic orchestration, Ultra is in a league of its own.

Hardware Requirements and License

Ultra is not a model for consumer hardware. Minimum deployment requires 4× NVIDIA B200 GPUs, 4× GB200/GB300, or 8× H100 GPUs — squarely in the data center tier. Pre-configured deployment recipes are available for vLLM, SGLang, and TRT-LLM.

On the licensing front, NeMoTron 3 Ultra is released under the OpenMDW-1.1 license — a permissive, commercially friendly agreement that permits commercial and non-commercial use, modification, distribution, and deployment in AI agents and services. NVIDIA has also released open pre-training datasets, post-training datasets, and full training recipes — a gold standard for open-weight AI transparency.

The Verdict

NVIDIA NeMoTron 3 Ultra is a specialized powerhouse rather than a general-purpose jack-of-all-trades. Its agentic capabilities are best-in-class, inference speed is remarkable, and open licensing makes it an attractive foundation for enterprise AI. But the inconsistency at pure code generation is a real limitation teams should factor in.

For organizations building autonomous agents, DevOps copilots, or complex workflow automation, Ultra is arguably the best open model available today. For those seeking a pure coding assistant, the search continues — but with a model this capable at everything else, the coding gaps will likely close soon.

For more details, visit the NVIDIA Build playground or check the official technical report.