Breaking

NVIDIA Releases Nemotron 3.5 Lightning for Long-Running AI Agents

September 9, 2026 at 16:12 EDT

Foundation Models
AI Agents
Open Source

NVIDIA has released Nemotron 3.5 Lightning, an open-weight 30B mixture-of-experts model designed for the high-volume execution layer of long-running AI agents. NVIDIA says it delivers up to 4x the output speed of similar-sized models and completes 10,000 PinchBench tasks 30% faster than Qwen3.6 35B at comparable accuracy.

August 11, 2026 · NVIDIA

Nemotron 3.5 Lightning targets agent execution

NVIDIA's open-weight 30B MoE model activates 3B parameters per token and is designed for fast, repeated work inside long-running AI agents.

30B

Total parameters

Active per token

Maximum output-speed claim

86%

PinchBench accuracy

The active compute is one-tenth of total capacity

Each block represents roughly 3B parameters. The model contains ten blocks in total, but only one is active for each token.

30B total

3B active

Takeaway: MoE routing keeps per-token compute to one of ten proportional blocks.

One system, different jobs

Plan
Frontier reasoning model

→

Route
Choose the right model

→

Execute
Lightning handles volume

What early tests liked

Fast local inference
Strong tool calling in agent workflows
Large context and customization options

Where caution remains

Code quality can lag expectations
Tool calling needs the correct parser flags
Speed varies by hardware and serving setup

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.