AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
AI models Introducing NVIDIA Nemotron 3.5 Lightning NVIDIA’s Nemotron 3.5 Lightning, now on Baseten, delivers high-throughput, efficient reasoning for faster and more accurate agentic workflows. Authors Marylise Tauzia Albert Lee Last updated August 11, 2026 Share TL;DR NVIDIA Nemotron 3.5 Lightning is a 30B MoE model with 3B active parameters and distilled from Nemotron 3 Ultra, built for agentic use cases. Nemotron 3.5 Lightning achieves leading accuracy on agentic tasks and also 4x higher throughput compared to other leading open models in its class and 30% lower time to task completion. Now available on Baseten Dedicated Inference, Lightning provides the high-throughput your always-on agents require. NVIDIA Nemotron 3.5 Lightning is now available on Baseten through Dedicated Inference, and it runs on NVIDIA accelerated infrastructure. Nemotron 3.5 Lightning is distilled from NVIDIA Nemotron 3 Ultra, a powerful model built for agentic tasks. Nemotron 3.5 Lightning inherits core agentic capabilities while optimizing for the efficiency, high throughput, and rapid token generation required for the specialized, high-volume agentic workflows highlighted below: Personal agents: Manage email, calendar, projects, and bookings. Financial services: Extract data, check policies, monitor risk, and summarize reports. Cybersecurity: Enrich alerts, classify incidents, query logs, and prepare findings. Telecom: Triage network alarms, optimize configurations, and answer billing questions. Retail: Enrich catalogs, resolve inventory exceptions, and assist product discovery. A small and speedy architecture for agents Nemotron 3.5 Lightning is a 30B Mixture-of-Experts (MoE) model with 3B active parameters. Designed specifically for always-on agents, this hybrid MoE architecture balances efficient processing with the reasoning power needed for complex, multi-turn tasks. It supports a 1M token context window, allowing agents to maintain deep context across long interactions. The model is lightning-fast; its 3B active parameters and multi-token prediction architecture enable almost 4 times higher throughput compared to other open models of similar size, and its rapid token generation capabilities allow agents to complete tasks more quickly. 30% Lower task completion time Nemotron 3.5 Lightning delivers performance on quality benchmarks comparable to other leading open models of similar size and can complete tasks 30% faster. ✕ PinchBench Accuracy vs. Time To Complete 10,000 Tasks Notably, Nemotron 3.5 Lightning’s speed advantage enables lower task-completion times through faster reasoning. By completing each reasoning step more quickly, agents can move from planning to execution sooner, iterate rapidly, and finish complex workflows in less time. Fine-tuning Lightning to improve quality Baseten worked with CodeRabbit, the AI code review application, to provide early access to Nemotron 3.5 Lightning on Baseten. The goal was to leverage the model to run a routing decision at the start of every code review. Nemotron 3.5 Lightning reads a code change and assigns complexity tags, and a scorer converts those tags into a review configuration. It's a very high-volume model call, which made it the right place to test whether a smaller model could carry a real piece of the pipeline. CodeRabbit post-trained Lightning in two stages, both on Baseten Training. The first was supervised fine-tuning (SFT), using NVIDIA NeMo AutoModel on managed H100 Training Jobs, which moved exact route agreement from 75.8% to 80.4% on a frozen 1,000-task evaluation. For the second they ran reinforcement learning with verifiable rewards through NVIDIA NeMo RL, scored against CodeRabbit's own routing policy. Cohen's kappa went from 0.461 to 0.544 and route accuracy held. ✕ SFT loss on managed H100 training jobs The finished adapter went straight to Dedicated Inference. It loads as a single rank-16 LoRA over stock Nemotron 3.5 Lightning and serves on a relatively small GPU like A100. Throughput measured 314.82 aggregate output tokens per second across eight concurrent requests. On the same workload the tuned model achieved ~4% higher accuracy than the baseline, ran at roughly half the cost of the baseline API calls, and produced 63.4% fewer output tokens getting there. Deploy Nemotron 3.5 Lightning on Baseten today Achieve higher accuracy through superior speed. Deploy Nemotron 3.5 Lightning on Baseten Dedicated Inference to power your autonomous agents with the fast reasoning they need to solve complex tasks. Subscribe to our newsletter Stay up to date on model performance, inference infrastructure, and more.