NVIDIA Nemotron 3.5 Lightning
by NVIDIA
Open 30B mixture-of-experts model tuned for the high-volume execution layer of long-running agents
Nemotron 3.5 Lightning is NVIDIA's open-weight 30B mixture-of-experts language model with only 3B active parameters per token, released on 11 August 2026 for the high-volume, latency-sensitive steps inside long-running AI agents. It ships with weights, training data and recipes under a permissive licence, alongside NeMo Switchyard, an open-source router that sends each step of an agent workflow to the cheapest model that can handle it.
NVIDIA released Nemotron 3.5 Lightning on 11 August 2026 as an open 30-billion-parameter mixture-of-experts model that activates only 3 billion parameters per token, targeting what NVIDIA calls the execution layer of long-running agents — the high-volume, repetitive tool calls and sub-tasks that dominate an agent's token spend. The architecture is a hybrid of Mamba-2 state-space layers, attention and MoE routing, supports context up to 1 million tokens, and covers English, Spanish, French, German, Italian and Japanese plus 43 programming languages. It was pre-trained on more than 20 trillion tokens with a September 2025 data cutoff, then post-trained with synthetic reasoning data, tool-calling datasets and GRPO reinforcement learning. NVIDIA reports 86% accuracy on PinchBench while completing 10,000 tasks 30% faster than Qwen3.6 35B at comparable accuracy, up to 4x the output speed of similar-sized models, 81.62% on MMLU Pro and 75.57% on GPQA Diamond. Weights, training data and recipes are published under the permissive OpenMDW-1.1 licence on Hugging Face and ModelScope, with NVFP4 and BF16 checkpoints and speculative-decoding draft models (DSpark, DFlash, MTP). It runs on vLLM, TensorRT-LLM, SGLang, Ollama, LM Studio and llama.cpp across Blackwell, Hopper and Ampere GPUs, from a Jetson or RTX 5090 up to DGX and the major clouds, and is served free on OpenRouter and as an NVIDIA NIM microservice on build.nvidia.com. The companion release, NeMo Switchyard, is an open-source model-routing library on GitHub offering tuning-free routers that pick a model per workflow step by capability, cost and latency.
A platform team running agents at high token volume that is paying frontier-model prices for repetitive execution steps that do not need frontier-model reasoning.
Frontier-adjacent tool-calling accuracy at roughly a tenth of the active parameters, with open weights that can be self-hosted or post-trained on private data.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Free, Contact for pricing
- Target Market
- CTOs, AI Platform Teams, ML Engineers, Enterprise Developers, Heads of AI
- Deployment
- Open-source, Self-hosted, API-based, Multi-cloud, Edge-first
- Founded
- 1993
- Headquarters
- Santa Clara, California, United States
- Team Size
- 500+
- Customers
- Named launch partners include Boomi, Cadence, Cognition, Kong, LangChain, LiteLLM, Nous Research, Ramp and Siemens
Key Features
- ✓30B mixture-of-experts with 3B active parameters
Sparse routing means only a tenth of the weights fire per token, cutting inference cost and latency while keeping a 30B model's breadth of knowledge.
- ✓Mamba-2 plus attention hybrid architecture
State-space layers handle long sequences at lower cost than pure attention, which is what makes the 1M-token context length economically usable in practice.
- ✓Harness-optimised tool calling
Post-training on tool-calling datasets with GRPO reinforcement learning targets accurate function-call arguments and reduced latency in agent harnesses specifically.
- ✓Fully open release under OpenMDW-1.1
Weights, training data and post-training recipes are all published with commercial use permitted, so teams can audit, self-host and re-train rather than rent.
- ✓NVFP4 and BF16 checkpoints with speculative decoding
Ships quantised for Blackwell alongside BF16, with DSpark, DFlash and MTP draft models for multi-token prediction, raising throughput on existing hardware.
- ✓NeMo Switchyard model router
Companion open-source library routes each agent workflow step to the best-fit model, with an LLM classifier, a stage router reading recent tool activity and a cost-escalation router.
- ✓Broad runtime and hardware coverage
Runs on vLLM, TensorRT-LLM, SGLang, Ollama, LM Studio and llama.cpp across Blackwell, Hopper and Ampere, from Jetson and RTX 5090 to DGX and major clouds.
Capabilities
Use Cases
- •Cost floor for agent execution steps
Route the repetitive tool calls and sub-tasks of a long-running agent to Lightning while reserving a frontier model for the genuinely hard planning steps.
- •Automated code review at volume
NVIDIA cites CodeRabbit with Baseten using Nemotron for review workloads where per-diff latency and cost matter more than maximum reasoning depth.
- •Security operations triage
CrowdStrike is named as a partner applying Nemotron models to cybersecurity workflows, where high alert volume makes per-token cost the binding constraint.
- •Domain post-training on private data
SiliconANGLE reports partner examples specialising the model for roughly $85 in about two hours on a single H100, using published NeMo recipes.
- •Local and on-device agents
Runs on a single RTX 5090, Jetson or DGX Spark via Ollama or llama.cpp, keeping sensitive prompts and documents entirely off third-party infrastructure.
Ideal For
Best For
- ✓The high-volume execution steps of long-running agents, where latency and cost per call dominate
- ✓Self-hosted or air-gapped deployments where open weights and a permissive licence are a hard requirement
- ✓Post-training on proprietary enterprise data using NVIDIA NeMo, with published recipes and training data
- ✓Long-context workloads up to 1M tokens, such as codebase-wide or document-set analysis
- ✓Local and edge inference on a single RTX 5090, DGX Spark or Jetson device
Not Ideal For
- ✗Hard multi-step reasoning and open-ended creative or complex coding work — Hacker News practitioners reported it going off the rails on a collaborative-whiteboard build where smaller dense models such as Qwen3.6-27B and Gemma 4-31B did better
- ✗Teams with no GPU operations capability who want a fully managed frontier model; getting value here means running vLLM, TensorRT-LLM or NIM yourself
- ✗Buyers who need vendor-backed accuracy guarantees or an SLA on model behaviour, since the open weights come as-is under OpenMDW-1.1
- ✗Workloads outside its six supported natural languages, where multilingual coverage is untested
Integrations
Deployment
Market & Ratings
Named launch partners include Boomi, Cadence, Cognition, Kong, LangChain, LiteLLM, Nous Research, Ramp and Siemens
Market Analysis
Pros
- ✓Genuinely open release — weights, training data and recipes under a permissive commercial licence — which is rarer than the word open usually implies
- ✓Strong measured throughput for its class, with 10,000 tasks completed 30% faster than Qwen3.6 35B at similar accuracy
- ✓Runs from a single consumer GPU to a datacentre, on every mainstream inference runtime, so it is cheap to trial
- ✓Post-training on private data is documented and inexpensive — partner examples cite roughly $85 on a single H100
- ✓Free to evaluate via OpenRouter and build.nvidia.com with no procurement step
Cons
- ✗Hacker News practitioners found it weak on complex, open-ended work: one reported it 'couldn't get the job done at all, went way off the rails' building a collaborative whiteboard, where smaller dense models did better
- ✗Commenters described over-thinking behaviour, with the model sketching several failed attempts in its reasoning trace before returning a poor answer
- ✗The MoE design was specifically criticised for struggling to hold two concerns at once — one commenter noted it could not write code and reason about a separate task simultaneously
- ✗SiliconANGLE notes there is no independent third-party validation of NVIDIA's performance claims, and the cited training-cost figures are anecdotal partner cases rather than systematic analysis
- ✗Hardware-vendor incentives are worth naming: NVIDIA's reason to publish open models is to sell GPUs, so the benchmark framing favours throughput on its own silicon
- ✗Data cutoff is September 2025 and multilingual coverage is limited to six natural languages
Pricing
Open weights (OpenMDW-1.1)
$0
- ✓Weights, training data and post-training recipes on Hugging Face and ModelScope
- ✓Commercial use permitted
- ✓NVFP4 and BF16 checkpoints
- ✓Self-host on vLLM, TensorRT-LLM, SGLang, Ollama or llama.cpp
OpenRouter hosted
$0
- ✓Free tier endpoint for evaluation
- ✓No local GPU required
NVIDIA NIM microservice / NVIDIA AI Enterprise
Contact for pricing
- ✓Packaged, supported container on build.nvidia.com
- ✓Free prototyping API credits
- ✓Enterprise support and lifecycle management
The model itself is free: weights, training data and recipes are published under OpenMDW-1.1 with commercial use permitted, and NeMo Switchyard is open source on GitHub, so the only cost of self-hosting is the GPUs you already run. A free hosted endpoint on OpenRouter and free prototyping credits on build.nvidia.com make evaluation zero-cost. The paid path is packaging and support, not the model — a production NIM microservice runs under NVIDIA AI Enterprise, whose per-GPU list pricing NVIDIA does not publish on the Nemotron pages and which requires a sales conversation. Third-party inference providers such as River AI resell it at $0.30 per million prompt tokens and $0.80 per million completion tokens.
Security & Compliance
Connect
Sources
This page was written from 7 sources, 6 on domains other than nvidia.com.
- 1.nvidia.com — nemotronvendor
- 2.blogs.nvidia.com — nemotron lightning switchyard rtx dgx
- 3.developer.nvidia.com — nvidia nemotron 3 5 lightning delivers fast accurate special
- 4.huggingface.co — NVIDIA Nemotron 3.5 Lightning 30B A3B NVFP4
- 5.siliconangle.com — nvidia releases nemotron 3 5 lightning nemo switchyard give
- 6.marktechpost.com — nvidia ai releases nemotron 3 5 lightning and nemo switchyar
- 7.datacamp.com — nemotron 3 5 lightning
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
River AI
Token-metered LoRA fine-tuning and reinforcement learning on open-weight models you keep
Scaled Cognition
APT, a large action model trained to take actions instead of predicting text
Engram
A learned memory layer that makes AI actually know your organization — at up to 100x fewer tokens.
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.