Skip to content

Nemotron 3.5 Lightning Runs Agent Tasks Up to 4x Faster

Original: Nemotron 3.5 Lightning Runs Agent Tasks Up to 4x Faster View original →

Read in other languages: 한국어日本語
LLM Aug 12, 2026 By Insights AI (Twitter) 2 min read 1 views Source
Nemotron 3.5 Lightning Runs Agent Tasks Up to 4x Faster

A small execution model for always-on agents

Long-running agents spend much of their budget on repetitive execution rather than difficult planning: calling tools, validating outputs, formatting results and delegating subtasks. NVIDIA Nemotron 3.5 Lightning is a 30B mixture-of-experts model built for that layer. It activates only 3B parameters for each token, aiming to combine the capacity of a larger model with the compute profile of a smaller one.

“An open 30B MoE model with 3B active parameters, built for always-on agents.”

In the source post, NVIDIA claims up to four times the output speed of similarly sized models. A follow-up post reports 86% accuracy over 10,000 PinchBench tasks, with a shorter completion time than Qwen3.6 35B at comparable accuracy. The distinction matters: useful agent throughput depends on completing tasks, not merely generating tokens quickly.

Weights, training data and a router

NVIDIA’s technical breakdown positions Lightning beside larger reasoning systems such as Nemotron 3 Ultra. A frontier model can handle orchestration and hard planning, while Lightning processes high-volume steps such as Git operations, tool-result checks and structured formatting. The newly released NeMo Switchyard library supplies a routing layer for sending each request to an appropriate open or closed model.

The release includes BF16 and NVFP4 checkpoints, multi-token prediction and draft models for speculative decoding. NVIDIA is also publishing weights, data and training recipes under its OpenMDW-1.1 terms. The accompanying Nemotron-RL Agentic Terminal Pivot dataset was used to post-train coding-agent capabilities. Deployment targets range from DGX Spark systems to data centers, with support across tools including Ollama, llama.cpp, LM Studio and Unsloth.

NVIDIA’s account typically covers GPU infrastructure, inference software and models tuned to its hardware ecosystem. Here, the broader move is architectural: model routing is becoming a first-class component, allowing expensive reasoning capacity to be reserved for the steps that need it.

What to test next

The “up to 4x” figure will vary with hardware, concurrency and decoding configuration. Independent tests should measure end-to-end task completion, error accumulation over long runs and the overhead introduced by routing. Developers should also check whether the 3B-active design preserves accuracy on their own tool schemas and domain data after customization.

Share: Long

Related Articles

LLM Reddit Mar 16, 2026 2 min read

A high-signal LocalLLaMA thread on March 15, 2026 focused on a license swap for NVIDIA’s Nemotron model family. Comparing the current NVIDIA Nemotron Model License with the older Open Model License shows why the community reacted: the old guardrail-termination clause and Trustworthy AI cross-reference are no longer present, while the newer text leans on a simpler NOTICE-style attribution structure.