FULL STORY

Nemotron 3.5 Lightning: From Launch to Fine-Tuned Results

NVIDIA launched the 30B open-source agent model Nemotron 3.5 Lightning alongside the NeMo Switchyard routing library. Subsequent fine-tuning by multiple labs showed performance rivaling Claude Opus.

2026-08-11 ~ 2026-08-21 · 8 episodes · 74 posts

Episode 1 · NVIDIA Unveils Nemotron 3.5 Lightning and NeMo Switchyard Router (2026-08-11, 55 posts)

NVIDIA released Nemotron 3.5 Lightning and NeMo Switchyard on August 11, targeting efficient execution for AI agents from edge to cloud. The model uses a hybrid Mamba-Transformer MoE architecture with 30B total parameters (31.6B measured) and 3B active (3.6B measured), supporting 1M context and measured 670 tokens/s output, leading in agentic coding tasks. Weights are open-sourced on Hugging Face, runnable on single GPU, with NVFP4 version topping HF trends. NeMo Switchyard dynamically routes based on request signals, agent state, tool results, and model capabilities, cost, latency.

Confirmed

  • Architecture and performance: hybrid Mamba-Transformer MoE, 30B total (31.6B measured), 3B active (3.6B measured), 1M context, NVFP4 and BF16, 670 tokens/s measured.
  • Open source and deployment: weights on Hugging Face, free, single GPU, DGX Spark demo, Baseten and Ollama support, NVFP4 topped HF trends.
  • Router: NeMo Switchyard open source, multi-dimensional dynamic routing.

Why it matters

Targets AI agent execution layer pain points, reducing high-frequency call cost and latency via ultra-low active parameter MoE, and with Switchyard dynamic routing, improves compute allocation for complex tasks, promoting always-on agents.

35 more related posts →

Episode 2 · Fine-tuned NVIDIA Nemotron 3.5 Outperforms Claude Opus (2026-08-12, 4 posts)

Following targeted post-training and fine-tuning by Harvey, Trajectory Labs, and Distil Labs, NVIDIA's Nemotron 3.5 Lightning outperformed Claude Opus on legal agent benchmarks and topped six different tasks while significantly reducing token consumption.

Episode 3 · Nvidia Switchyard Router Cuts Agent Costs by 59% (2026-08-12, 3 posts)

Nvidia introduced the Switchyard router to dynamically switch AI models mid-task, significantly reducing operational costs. Ramp's tests on coding agents show a 59% cost reduction and 32% faster execution times without compromising performance.

Episode 4 · NVIDIA Nemotron 3.5 Excels in Scientific Post-Training (2026-08-12, 2 posts)

Lila Sciences reported that the NVIDIA Nemotron 3.5 Lightning preview excels in post-training for scientific tasks, mastering material science reinforcement learning environments faster and easier than previous models.

Episode 5 · Meta Open-Sources 30B Local Agent Model Muse Glimmer (2026-08-12, 2 posts)

Meta's Super Intelligence Lab has open-sourced Muse Glimmer, a 30B multimodal model designed for local agents. Available on Ollama with an Apache 2.0 license, it requires only 24GB of VRAM to run locally.

Episode 6 · Meta and NVIDIA Launch 30B Open-Source Agent Models on the Same Day (2026-08-12, 3 posts)

Meta and NVIDIA each released 30B parameter open-source agent models on the same day: Muse Glimmer and Nemotron 3.5 Lightning. Both focus on strong tool-calling and cost-effective routing, presenting a choice between local agent execution and cloud-based routing.

Episode 7 · NVIDIA Open Sources NeMo Switchyard for Dynamic Agent Routing (2026-08-15, 2 posts)

NVIDIA released the open-source NeMo Switchyard library, enabling AI agents to dynamically route between models based on task complexity.

Episode 8 · NVIDIA Unveils Nemotron 3.5 Lightning and NeMo Switchyard (2026-08-19, 3 posts)

NVIDIA released the 30B-parameter MoE model Nemotron 3.5 Lightning for long-running agent workflows, plus the open-source NeMo Switchyard library that routes each agent workflow step to the best model based on quality, latency, and cost.