Editorial archive
Local
← Back to home
0005

Record

Brief · Infrastructure

Nemotron 3.5 Lightning and the Routing Layer

A purpose-built MoE executor that activates only 3 billion parameters per token, paired with NeMo Switchyard routing, is redrawing the cost structure of agentic workloads — though the speed claims remain vendor-reported for now.

Contents
  1. 01 · A Purpose-Built Executor That Wakes 3 Billion Parameters per Token
  2. 02 · A Routing Layer That Segments Requests by Cost Tier
  3. 03 · 4x Speed Claims and a Verification Gap
  4. 04 · Sources

A Purpose-Built Executor That Wakes 3 Billion Parameters per Token

Released as an open model, Nemotron 3.5 Lightning features a hybrid mixture-of-experts (MoE) architecture that activates 3 billion of its 30 billion total parameters per token ¹³. Distilled from Nemotron 3 Ultra, the model was developed in collaboration with the Nemotron Coalition, whose members contributed evaluation methodologies, datasets, and inference software ¹³. It supports context windows of up to 1 million tokens and ships in BF16 and NVFP4 checkpoints, with deployment targets that NVIDIA says span edge devices like Jetson through to H100 and B300 datacenter GPUs ³. Open-source inference engine vLLM announced day-0 support, serving the model through an OpenAI-compatible API ³.

A Routing Layer That Segments Requests by Cost Tier

NeMo Switchyard, released alongside the model, is an open-source routing library that runs directly inside popular agent tools. It intelligently directs each request to the most suitable model across a developer’s own mix of open, proprietary, and NVIDIA models — without requiring application rewrites ¹². The design rests on the idea that agent architectures are evolving into multi-model systems divided by role: frontier reasoning models such as Nemotron 3 Ultra or GPT-5.6 plan and orchestrate workflows, while compact executors like Lightning handle numerous, well-scoped steps — including code review, tool use, security alert monitoring, and billing questions ¹². NVIDIA says companies including CrowdStrike, Harvey, and CodeRabbit are already customizing the model for their own domains ¹.

4x Speed Claims and a Verification Gap

NVIDIA reports up to 4x higher output speed than similarly sized models, achieved through speculative decoding, multi-token prediction, and NVFP4/BF16 quantization, and claims this translates into 30% faster agentic task completion versus models in its class ¹². Those figures come directly from the vendor’s own PinchBench measurements. Independent third-party latency and quality benchmarks for long-running agent workloads have not yet been published, and the hardware conditions behind the 4x figure remain to be verified externally ¹².

Sources

Next in reading

750 Tokens per Second at the Frontier: How Cerebras Wafer-Scale Powers GPT-5.6 Sol Ultrafast

OpenAI and Cerebras have unveiled Ultrafast, a tier running GPT-5.6 Sol at up to 750 tokens per second. Here is how wafer-scale design sidesteps the GPU memory bottleneck — and what the numbers still need to prove.

Read piece →