NVIDIA’s recently released open-source AI model “Nemotron 3 Super” has officially taken the top spot on the EnterpriseOps-Gym leaderboard, scoring an average of 27.3 points and outperforming numerous competitors including MiniMax M27, Kimi K2.5, DeepSeek v3.2, and GPT-OSS-120B. This once again proves that NVIDIA is not only the leading AI hardware manufacturer but also possesses top-tier capabilities in software and model development.
Nemotron 3 Super: A Hybrid Architecture Model Built for AI Agents
NVIDIA officially unveiled Nemotron 3 Super at the GTC conference this past March. This is a Mixture-of-Experts (MoE) model with 120 billion (120B) total parameters but only 12 billion (12B) active parameters. The model adopts a Hybrid Mamba-Transformer architecture and introduces several breakthrough technologies:
- Latent MoE: Compress tokens before they reach the expert layer, enabling the model to invoke up to 4x more experts at the same inference cost, achieving finer-grained specialization.
- Multi-token prediction (Multi-Token Prediction, MTP): A single forward pass can predict multiple future tokens, significantly reducing the generation time for long sequences, with built-in speculative decoding capability.
- Hybrid Mamba-Transformer Backbone: Mamba layers handle efficient sequence processing, Transformer layers maintain precise associative retrieval capabilities, and combining both achieves a 4x improvement in memory and computational efficiency.
- Native NVFP4 pre-training: Optimized specifically for the NVIDIA Blackwell architecture, delivering 4x faster inference on B200 compared to H100 using FP8 format, while maintaining model accuracy.
- Multi-Environment Reinforcement Learning: Post-training was conducted across 21 environment configurations using NVIDIA NeMo Gym and NeMo RL, with over 1.2 million rollouts trained.
Additionally, Nemotron 3 Super supports a native 1M-token context window, enabling AI Agents to maintain long-term memory during extended tasks and prevent goal drift, with throughput improved by over 5x compared to the previous generation Nemotron Super.
EnterpriseOps-Gym Real Test: 1,150 Tasks Validate Capabilities
EnterpriseOps-Gym is a benchmark designed to evaluate AI agents’ performance in enterprise environments, covering 1,150 tasks and featuring 512 functional tools. Models must coordinate across multiple enterprise systems and tools to complete a single workflow in a fully interactive environment, making it one of the most challenging enterprise-level AI benchmarks to date.
atOn the open-source model leaderboardNVIDIA Nemotron 3 Super claimed the top spot with an average score of 27.3, leading in TEAMS, Email, and Hybrid workflows, while also remaining highly competitive in CSM, ITSM, and Drive workflows. The specific rankings are as follows:
- NVIDIA Nemotron 3 Super — 27.3 points
- Kimi K2.5 — Second Place
- DeepSeek v3.2 — Third Place
- GPT-OSS-120B — 5th place
Third-Party Review: The Perfect Balance of Openness and Intelligence
According to independent benchmarking organization Artificial Analysis’sReportNemotron 3 Super scored 36 points on the intelligence index, 17 points higher than the previous generation Nemotron Super and outperforming GPT-OSS-120B’s score of 33. It achieved 83 points on the openness index, ranking second only to institutions such as Ai2 and MBZUAI, making it the most open and most intelligent open-source model.
The evaluation also noted that Nemotron 3 Super performs especially well on Agent tasks: it achieved a 29% score in Terminal-Bench Hard and reached 1027 ELO in GDPval-AA (which evaluates Agent performance in real-world work tasks). Additionally, despite having a relatively smaller parameter scale, its inference efficiency is extremely high, with per-GPU throughput approximately 10% higher than GPT-OSS-120B.
On PinchBench, a benchmark that evaluates LLMs as AI Agent brains, Nemotron 3 Super also achieved a score of 85.6%, making it the best-performing open model in its class.
Open ecosystem and price advantage
Nemotron 3 Super adopts a fully open license, not only opening up the weights but also publishing the training dataset and complete training recipes, allowing developers to freely customize, optimize, and deploy on their own infrastructure. The model is available on Hugging Face under the model ID “nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8”.
In terms of pricing, Nemotron 3 Super offers highly competitive inference costs, with input tokens at approximately $0.30 per million (around NT$9.8) and output tokens at approximately $0.80 per million (around NT$26). Through platforms like DeepInfra, prices can drop even further to just $0.10/$0.50 per million tokens.
The Nemotron 3 series currently includesthree modelsNano (30 billion total parameters, 3 billion active parameters), Super (120 billion total parameters, 12 billion active parameters), and the upcoming Ultra (approximately 500 billion total parameters, 50 billion active parameters). Combined with the recently announced Nemotron 3 Nano Omni (claimed to boost AI Agent throughput by 9x), NVIDIA is gradually building a complete open AI model ecosystem.
Conclusion
NVIDIA Nemotron 3 Super takes the top spot on the open-source model leaderboard, showcasing not only the company’s deep expertise in AI software but also proving that the Hybrid Mamba-Transformer combined with a latent MoE architecture can indeed achieve the optimal balance between efficiency and accuracy. As AI Agent applications become increasingly prevalent, models designed specifically for multi-step reasoning and tool collaboration will play an increasingly critical role in enterprise AI deployments.
Source: KOCPC Chinese