NVIDIA has released the open source model Nemotron 3.5 Lightning, specifically designed for long-duration Agent execution tasks. The model has a total of 30 billion parameters, with approximately 3 billion parameters activated for each Token, primarily responsible for high-frequency repetitive tasks such as tool invocation, result checking, and sub-Agent scheduling.

NVIDIA's design philosophy is to thoroughly divide the roles of Agents: complex planning is assigned to larger models like Nemotron 3 Ultra, while a large volume of repetitive execution is handled by Lightning. The official claim states that its output speed can reach up to 4 times that of similar models; it achieves an accuracy of 86% on PinchBench, and completes 10,000 tasks 30% faster than Qwen3.6 35B under similar accuracy conditions.

The model employs a MoE architecture and is specifically trained for Agent Harness, incorporating multi-Token prediction and speculative decoding. Officially provided BF16 and NVFP4 weights can run on local devices such as RTX 5090 and DGX Spark, and it supports llama.cpp, Ollama, LM Studio, and Unsloth.

NVIDIA has opened weights, some training data, and training recipes, allowing enterprises to continue post-training for tasks related to code, security, and legal matters. In the official demonstration cases, Lightning significantly improved scores on multiple specialized tasks after post-training.

From a market mechanism perspective, the efficient execution layer model reduces the operational costs and latency of long-duration Agents, concentrating funding and deployment on locally runnable, customizable open source execution models, while solutions that rely solely on large models to handle all tasks face pressure.

This model is positioned as a high-speed workhorse in multi-Agent systems, rather than a general model pursuing the highest intelligence.

Source: Public Information