NVIDIA Ships Nemotron 3.5 Lightning and NeMo Switchyard
On August 11, 2026, NVIDIA released Nemotron 3.5 Lightning—a 30B open MoE model with ~3B active parameters for high-volume agent execution—and NeMo Switchyard, an open-source model router that sends each agent step to the cheapest capable model, claiming up to 4× throughput in class and large cost cuts when paired with frontier planners.
TLDR
NVIDIA on August 11, 2026 expanded the Nemotron 3 open family with Nemotron 3.5 Lightning—a 30-billion-parameter mixture-of-experts model with roughly 3 billion active parameters, distilled from Nemotron 3 Ultra for always-on agent execution workloads—and NeMo Switchyard, an open-source model routing library that steers each request to the right model in a developer-defined pool. NVIDIA claims ~4× throughput versus comparable in-class models at similar intelligence and up to ~30% faster completion of agentic benchmark workloads. Weights land on Hugging Face / ModelScope / OpenRouter and as NIM microservices; Switchyard is on GitHub with gateway partners including OpenRouter, LiteLLM, and Kong.
What shipped
| Piece | Detail (NVIDIA blog / analyst notes) |
|---|---|
| Nemotron 3.5 Lightning | 30B MoE, ~3B active; hybrid Mamba-Transformer; multi-token prediction; latent MoE; speculative decoding |
| Role | High-volume specialized calls in long-running agents (tool calls, validation, formatting)—not frontier planning |
| Open stack | Open weights + post-training datasets + training recipes (Nemotron Coalition contributions) |
| Hardware targets | Jetson, GeForce RTX, DGX Spark, DGX Station; single-GPU laptop/desktop class claims |
| NeMo Switchyard | Open router: custom algorithms/policies across a model pool; mid-task reshuffle |
| Partner claim | LangChain-cited path: large cost reduction routing Lightning vs Claude Opus-class with small accuracy tradeoff (vendor/partner measured—not independent AA) |
| Availability | HF, ModelScope, OpenRouter, build.nvidia.com NIM; Switchyard on GitHub |
Ranking context
Lab / partner: Throughput and agent-workload latency claims are NVIDIA’s; treat as vendor benches with disclosed architecture advantages (MoE + MTP + speculative decoding).
Independent:
- Artificial Analysis composite intelligence (cited in contemporaneous analysis): Lightning around ~24 on AA’s index—near the bottom of the frontier chart and below NVIDIA’s own Nemotron 3 Super (~26), and trailing Gemma 4 31B—consistent with NVIDIA’s pitch that most agent tokens are execution, not reasoning.
- Do not present Lightning as a Sol/Opus peer; the product is cheap reliable worker nodes behind a router.
Product-line placement
Distinct from the $500B Wall Street financing MOUs (Aug 10) and Firebird factory open (Aug 8). Lightning + Switchyard is the open model + agent infrastructure software product line—NVIDIA selling the orchestration layer, not only GPUs.
Why this story matters
If agents spend most tokens on boring steps, the winning stack is frontier planner + fast open worker + router. Switchyard is strategically heavier than another small open model: it positions NVIDIA inside the decision of which model runs. Watch: real production cost curves; whether Switchyard becomes a default in LiteLLM/Kong gateways; and competition from Meta Glimmer-class local agents on the same laptops.