APREX · EST. 2026
With 550 billion total parameters and a hybrid architecture combining Mamba and Transformer, NVIDIA's Nemotron 3 Ultra is positioning itself as a serious contender in the increasingly competitive market for open large language models. The model was released on June 4, 2026 under the OpenMDW-1.1 license and is built specifically for agent-based workflows requiring long context and multi-step reasoning.
LangChain adapts Deep Agents for Nemotron
One of the most significant developments is that LangChain — one of the most widely used platforms for AI agent orchestration — has optimized its Deep Agents framework directly for Nemotron 3 Ultra. According to the NVIDIA blog, this combination achieves the highest accuracy among open models, with greater throughput and a price tag reportedly ten times lower than leading closed models.
Nemotron 3 Ultra promises to cut costs for agent tasks by up to 30 percent, while delivering the industry's highest speed among open models in its class.
Impressive numbers — with caveats
NVIDIA cites a range of benchmark figures worth examining more closely. The company claims the model delivers over 300 tokens per second on a pre-release DeepInfra endpoint, compared to an estimated 50–100 tokens per second for rival Kimi K2.6 via commercial APIs. That is a significant difference, though it should be noted that the figures for Kimi K2.6 are estimates and are not as precisely documented.
On the Artificial Analysis Intelligence Index, Nemotron 3 Ultra scores 48, making it the strongest American open-weight model to date. However, Kimi K2.6 from Moonshot AI leads with a score of 54 on the same index — a gap that is far from negligible.
Strong on context and instruction following
Where Nemotron 3 Ultra truly stands out is in long-context handling. The model scores 95 percent on the RULER benchmark with 1 million tokens of context, according to NVIDIA. Most competing open models top out at 256,000 tokens. The model supports up to 1 million tokens of context when using NVFP4 quantization on Blackwell GPU architecture.
On IFBench, which measures instruction following, Ultra scores 82 percent — higher than GLM-5.1 (77 percent), Kimi K2.6 (74 percent), and Qwen 3.5 (78 percent).
Weaker on coding and long-horizon planning
The picture is not entirely positive. On Terminal-Bench 2.0, which tests coding skills, Nemotron 3 Ultra scores 54 percent — behind Kimi K2.6's 67 percent and GLM-5.1's 64 percent. On EnterpriseOps-Gym, which measures long-horizon planning, the model scores 33 percent, well behind GLM-5.1's 40 percent.
SWEBench Verified — an industry-recognized benchmark for coding tasks in real-world projects — gives Ultra between 65 and 70.4 percent, depending on the agent framework. That is solid, but not exceptional.
Open, but not without competition
Nemotron 3 Ultra was trained on 20 trillion tokens and supplemented with 212 billion domain-specific tokens. NVIDIA has also released 10 million new SFT training samples and 1 million new reinforcement learning tasks to the research community.
Overall, Nemotron 3 Ultra emerges as a strong candidate for enterprises seeking high inference speed and low operational costs in agent-based systems — particularly within NVIDIA's own hardware ecosystem. But for developers who prioritize raw coding performance or overall intelligence scores, the benchmarks suggest that Chinese models such as Kimi K2.6 still hold the top spot on the podium.
Source material: NVIDIA AI Blog and independent benchmark analysis from Artificial Analysis.