← Back to feed

NVIDIA has released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model with 3 billion active parameters, targeting the execution layer of enterprise agentic stacks. The model ships in NVFP4 format on Hugging Face and claims 4x output speed and 30% faster agentic task completion against comparable open models.

The release pairs with NeMo Switchyard, NVIDIA’s new agent orchestration framework, and arrives under the Nemotron Coalition label — a formal programme in which external partners contributed evaluation methodologies, inference software, and training datasets to the model development process.

The Execution Layer Thesis

Nemotron 3.5 Lightning is not positioned as a frontier reasoning model. NVIDIA is explicit: it is the always-on execution layer that runs continuously beneath higher-capability models in a multi-agent stack.

The production pattern it targets looks like this: a frontier model (Fable 5, GPT-5.6 Sol, or Grok 4.5) plans and delegates. Nemotron 3.5 Lightning executes — running tool calls, parsing structured outputs, formatting data, completing repetitive subtasks — at a cost and latency profile that makes continuous background operation economically viable. Frontier model calls stay bounded to the decisions that require frontier capability.

This split-layer architecture has already emerged in production codebases. Claude Dynamic Workflows and AWS AgentCore both use it. NVIDIA is betting that as this pattern standardises, enterprises will want an open, auditable, customisable model for the execution tier rather than another proprietary API.

Architecture: Why 30B/3B Matters

The 30B total / 3B active MoE configuration is a deliberate point on the compute curve. Inference cost is determined by active parameters — the 3B that fire per token. Total capacity (30B) sets the knowledge ceiling and enables specialisation across expert modules without paying the forward-pass cost of a dense 30B model.

For agent workloads, the efficiency advantage compounds. Agent inference is dominated by generation at scale: formatting tool call responses, emitting structured JSON, producing short confirmations across thousands of concurrent threads. At 3B active parameters, Nemotron 3.5 Lightning can sustain that throughput on a single GPU without batching penalties.

The NVFP4 format — NVIDIA’s 4-bit floating-point representation optimised for Blackwell Tensor Memory Accelerators — extends the efficiency advantage on current-generation hardware. GB200 and B200 GPUs natively accelerate NVFP4 operations, which is why Lightning is distributed in that format rather than the more common BF16 or FP8 builds.

The Nemotron Coalition

External partners contributed evaluation methodologies, inference software, and datasets to the project. This structure differs from a standard model release. Coalition members essentially co-designed what “good” looks like for enterprise agent deployment — specifying benchmark tasks, writing test harnesses, and supplying domain-specific training data.

The practical result is that Nemotron 3.5 Lightning arrives pre-integrated with NemoClaw, NVIDIA’s open-source security and management stack for agent deployment, and with explicit documentation for OpenClaw and Hermes Agent harnesses. The model isn’t just weights on Hugging Face — it comes with a production deployment path.

NeMo Switchyard, launching alongside, handles the routing layer: deciding when a task should stay on Lightning versus escalate to a frontier model, managing tool call dispatch, and maintaining execution state across multi-step agent runs.

What Changes for Enterprise Deployments

The combination of Lightning and Switchyard gives infrastructure teams something they haven’t had from NVIDIA before: a complete open stack from chip to orchestration for agentic workloads, without proprietary API dependency.

The benchmark claims — 4x output speed, 30% faster task completion — need independent verification against a stated baseline, which NVIDIA has not clearly specified. What is verifiable: Lightning uses the same 30B/3B MoE form factor as Nemotron 3 Nano Omni, which reached 1,200+ tokens/second on Blackwell in early benchmarks. At that throughput, a single GB200 can sustain hundreds of concurrent agent threads on the execution tier while keeping a handful of frontier model slots for planning.

That economics is the actual case for this model. At frontier pricing ($3-15 per million output tokens), running planning-level models on execution tasks is prohibitively expensive at scale. Nemotron 3.5 Lightning, deployed on-premises on NVIDIA hardware, converts execution tokens from a margin item to an infrastructure cost.

Where It Fits in the Leaderboard

Artificial Analysis is running independent benchmarks on Nemotron 3.5 Lightning and published preliminary analysis today. Early positioning puts it in the 73-75 range on SWE-bench Verified — below the frontier tier (Claude Opus 5 at 97%, GPT-5.6 Sol at 96.2%) but competitive with models like DeepSeek V3 at 73% and above Gemini 2.5 Pro at 63.8%, while operating at a fraction of the inference cost.

The SWE-bench score is instructive because software engineering tasks are compositionally similar to agentic execution work: structured tool use, file manipulation, command output parsing. A mid-tier SWE-bench score at 4x output speed is the correct trade-off for the execution layer role NVIDIA is targeting.