NVIDIA released Nemotron 3.5 Lightning on August 11, 2026, an open-weight AI model the company says anyone can download, run, and modify without paying a license fee or asking permission. CNBC framed the release as Nvidia’s first open-source model launch since CEO Jensen Huang made open weights a bigger part of his public pitch to developers. The model is small enough to run on a single high-end GPU, yet NVIDIA is positioning it as the backbone for a very specific job: keeping AI agents running continuously without burning through a frontier model’s compute budget on every step.
That framing matters more than the spec sheet. A free, lightweight model that pushes more workloads onto GPUs, rather than away from them, tells you something about where Nvidia thinks the next leg of AI demand comes from. This piece breaks down what Nemotron 3.5 Lightning is, how it compares to its predecessor and at least one open-weight rival, where it can run, and what the release signals about Nvidia’s broader GPU strategy heading into the back half of 2026.
What Nemotron 3.5 Lightning Actually Is
Nemotron 3.5 Lightning is a mixture-of-experts (MoE) language model with roughly 30 billion total parameters, of which only about 3 billion activate for any given token. NVIDIA’s own developer blog post describes it as “an open 30B mixture-of-experts (MoE) model with 3B active parameters built for that execution layer of always-on agents.” Independent benchmarking site Artificial Analysis lists slightly more precise figures: 31.6 billion total parameters and 3.6 billion active per token, framing Lightning as the direct successor to Nemotron 3 Nano 30B A3B.
The gap between NVIDIA’s rounded marketing numbers and the more granular third-party count is normal for MoE releases, where exact expert and layer counts get simplified for a blog headline. What both sets of numbers agree on is the ratio: roughly one-tenth of the parameters are active at once, which is the entire point of an MoE design built for cheap, high-volume inference rather than peak reasoning depth.
Inside the Architecture: MoE, Mamba-2, and a 1M-Token Window
Technical write-ups describe Lightning as a hybrid design that blends Mamba-2 state-space layers with mixture-of-experts routing and attention, rather than a pure transformer stack. That hybrid approach is common among 2026-era efficiency models because Mamba-2 layers process long sequences with less memory overhead than standard attention, while MoE routing keeps the compute cost per token low.
Lightning ships with a context window up to 1 million tokens, which is large enough to hold an entire codebase, a long agent transcript, or a full support ticket history in a single pass. NVIDIA also shipped the model in two checkpoint formats: standard BF16 and the more compact NVFP4 quantization, alongside a pair of smaller draft models, DFlash and DSpark, used for speculative decoding. Speculative decoding lets a small draft model propose several tokens at once for the larger model to verify, which is how NVIDIA gets to its headline speed claims without shrinking the base model further.
Nemotron 3.5 Lightning at a Glance
| Attribute | Detail |
|---|---|
| Release date | August 11, 2026 |
| Architecture | Hybrid Mamba-2 + mixture-of-experts (MoE) + attention |
| Total / active parameters | ~30B total / ~3B active (31.6B / 3.6B per Artificial Analysis) |
| Context window | Up to 1,000,000 tokens |
| License | OpenMDW-1.1 (permissive, commercial use allowed) |
| Checkpoint formats | BF16, NVFP4, plus DFlash and DSpark draft models |
| Predecessor model | Nemotron 3 Nano 30B A3B |
| Primary use case | Execution layer for always-on autonomous agents |
| Minimum practical hardware | Single high-end GPU (e.g. RTX 5090 or H100) |
| Distribution channels | Hugging Face, build.nvidia.com NIM, Ollama, OpenRouter |
The OpenMDW-1.1 License: What “Free” Actually Means
NVIDIA shipped Lightning under the OpenMDW-1.1 license, described by Artificial Analysis as open for commercial use without material restrictions. CNBC reported that the model is free for companies to download, use, and modify without getting permission or paying Nvidia anything. That is a meaningfully more permissive stance than a research-only or non-commercial license, and it puts Lightning in the same competitive bucket as other permissively licensed open-weight releases from Qwen, DeepSeek, and Meta’s Llama line.
The practical upshot: a startup or enterprise team can fine-tune Lightning, embed it in a paid product, and never write NVIDIA a check for the privilege. NVIDIA does supply its own fine-tuning tooling for teams that want it, including LoRA and full supervised fine-tuning through NeMo Automodel, plus reinforcement learning support via NeMo RL and NeMo Gym, alongside a purpose-built Nemotron-RL Agentic Terminal Pivot dataset for training agent behaviors.
Benchmark Numbers: How Lightning Stacks Up
NVIDIA’s own claim, repeated across its developer blog and its X account, is that Lightning delivers up to 4x the output speed of similarly sized models. On accuracy, Artificial Analysis places Lightning at 24 on its Intelligence Index, a composite benchmark score the site uses to rank models. That is a 9-point jump over Nemotron 3 Nano’s score of 15, which Artificial Analysis calls a significant intelligence improvement generation over generation.
On task-specific benchmarks reported for the BF16 checkpoint, Lightning scores 51.56 on SWE-bench Verified (a widely used coding-agent benchmark), 85.37 on PinchBench, 75.44 on GPQA Diamond, and 81.94 on MMLU Pro. None of those numbers put Lightning at the top of the open-model leaderboard for raw reasoning. That is by design: NVIDIA is not chasing an accuracy crown here, it is optimizing for throughput on repetitive, well-defined tasks where a bigger, slower model would be overkill.
Nemotron 3.5 Lightning vs. Nemotron 3 Nano: The Generational Jump
Lightning replaces Nemotron 3 Nano 30B A3B as the smallest model in NVIDIA’s current Nemotron 3 family, and the generational comparison is where the release looks most convincing. A 9-point Intelligence Index gain (15 to 24) at a similar parameter footprint means NVIDIA squeezed meaningfully more capability out of roughly the same compute envelope, largely through the architecture change (adding Mamba-2 layers) rather than simply scaling parameters up.
That pattern (holding size flat, improving quality through architecture and training data) tracks with a broader shift across the open-model field this year. NVIDIA has also accelerated its own release cadence to roughly every four to six weeks, which puts pressure on every generation to show a clear improvement over the last one or risk looking stale within a single quarter.
Nemotron 3.5 Lightning vs. Qwen3.6-35B: The Open-Weight Race
NVIDIA and its ecosystem partners have also drawn a direct comparison against Alibaba’s Qwen3.6-35B, a similarly sized open-weight competitor. On a 10,000-task run of PinchBench, Lightning reportedly completed the workload 30% faster than Qwen3.6-35B while holding accuracy in a similar range, around 86%. That is a narrower, more concrete claim than the general 4x speed figure, and it is the kind of head-to-head number that matters most to a team picking between two similarly priced open models for a production agent pipeline.
The Qwen family has been moving fast in the same direction. Shattered.io covered Qwen3.8-Max’s open-weight release scoring 86.6 on its benchmark suite, edging closer to closed frontier models. Lightning is not trying to compete in that tier. It is aimed at a cheaper, faster lane one level below, where speed and cost per token matter more than topping a leaderboard.
Where It Runs: From a Single RTX 5090 to DGX Spark and the Cloud
The headline consumer-facing claim, per CNBC, is that Lightning is lightweight enough to run on a single GPU on a laptop or desktop. That is a real shift from frontier-class models that require multi-GPU clusters just to load. At the other end of the spectrum, Oracle Cloud Infrastructure’s release notes list specific minimum GPU shapes for importing Lightning into its Generative AI service, spanning A100, H100, H200, and B200 configurations, and Amazon made the model available through Amazon SageMaker JumpStart.
| Where it runs | Access path | Notes |
|---|---|---|
| Single RTX GPU (desktop/laptop) | Ollama, LM Studio, llama.cpp, Unsloth | Local inference, no cloud bill |
| NVIDIA DGX Spark / DGX Station | NVIDIA NIM via build.nvidia.com | Local “AI workstation” tier |
| H100 / H200 / B200 / A100 (cloud) | AWS SageMaker JumpStart, Oracle OCI Generative AI | Enterprise data-center shapes |
| Serverless inference | DeepInfra, Fireworks, FriendliAI, CoreWeave, GMI Cloud, Nebius, Crusoe | Pay-per-token hosted access |
| Open weights, self-hosted anywhere | Hugging Face, ModelScope, OpenRouter | No NVIDIA hardware requirement to download |
For developers who want to try Lightning locally first, the Ollama listing makes it a one-line pull once the tag is published to the library:
ollama pull nemotron
ollama run nemotronThat kind of frictionless local access, paired with data-center-grade deployment paths for teams that need to scale, is the through-line of the release. NVIDIA wants the same model available whether a developer is testing an agent idea on a gaming GPU or running it in production across a fleet of H200s, and it is not charging a toll at any point on that path. Readers running local inference on their own hardware can compare notes against our earlier guide to running open LLMs locally on an RTX 5090.
Built for Always-On Agents, Not Chatbots
NVIDIA is explicit that Lightning is not meant to be a general-purpose chatbot. The developer blog post’s title calls it a model for “specialized task execution for long-running agents,” and the company’s X post describes it as “built for always-on agents to complete high-volume, specialized tasks faster.” NVIDIA also said: “Nemotron 3.5 Lightning is a customizable open 30B MoE model with 3B active parameters, providing optimal high-volume execution for autonomous agents,” in the same launch post on its developer blog.
NVIDIA lists compatibility with agent harnesses including OpenClaw and Hermes Agent, and points to NemoClaw, an open-source security and management stack, as the layer that handles governance for agents built on Lightning. The pitch is a two-tier stack: a larger, slower reasoning model plans and makes judgment calls, while Lightning executes the repetitive steps (tool calls, retrieval lookups, structured outputs) at a fraction of the cost per token.
The GPU Demand Angle: Why Give Away a 30B Model?
The financial logic behind a free model is not charity. Industry analysis has tied the Lightning release to Nvidia’s data-center business, which reports have pegged at roughly $89 billion in revenue, arguing that a fast, free execution-layer model expands the pool of workloads tuned to run well on NVIDIA hardware. Every always-on agent built on Lightning is a workload that runs continuously rather than in short chat bursts, and continuous workloads consume more aggregate GPU-hours than sporadic ones, even when the model itself carries no license fee.
Shattered.io tracked a related thread in Nvidia’s record $96.2 billion quarterly results, where data-center GPU demand was the dominant story even as rivals circled. Giving away a model that is explicitly optimized to run best on NVIDIA silicon, and that is designed to keep GPUs busy around the clock, fits squarely into that same demand-generation strategy rather than working against it.
Historical Context: NVIDIA’s Open-Model Strategy
NVIDIA has been building out the Nemotron family for several release cycles, moving from general-purpose reasoning variants toward a more segmented lineup that separates planning-tier models from lightweight execution-tier models like Lightning. That segmentation mirrors a pattern playing out across the industry in 2026: rather than one model trying to do everything, vendors increasingly ship a family, with a flagship for hard reasoning and a cheaper, faster sibling for volume work.
NVIDIA’s broader AI ecosystem push has included its reported multi-billion-dollar interest in Hugging Face, the platform that now also hosts Lightning’s open weights. Whatever the outcome of that deal, distributing Lightning through Hugging Face keeps it in front of the same developer audience NVIDIA would be paying to reach through any acquisition, reinforcing that open distribution is now a core part of the company’s go-to-market playbook rather than a side project.
Market Impact: Revenue, Rivals, and the Open-Weight Arms Race
Business coverage of the launch has generally treated it as incremental positive news rather than a surprise. CNBC framed it as part of Nvidia’s push to broaden AI accessibility under Huang’s more open public messaging, without reporting a sharp stock move tied specifically to the announcement. That is consistent with how the market has treated prior open-weight releases from major labs this year: useful for competitive positioning and developer goodwill, but not a standalone catalyst the way an earnings beat or a large cloud contract would be.
Where the release does matter commercially is in the crowded field of serverless inference providers now hosting Lightning at launch, including DeepInfra, Fireworks, FriendliAI, CoreWeave, GMI Cloud, Nebius, and Crusoe. That list signals real demand from hosting providers to have a fast, cheap, permissively licensed model on their menus fast, and it gives smaller GPU cloud players a way to compete on price for agent workloads without having to train a comparable model themselves.
Competitive Landscape: NVIDIA vs. the Open-Model Field
Lightning enters a field that already includes Qwen’s 30-40B class models, DeepSeek’s open releases, and Meta’s Llama line, all competing on a similar axis of size versus speed versus accuracy. NVIDIA’s edge is not that Lightning is the smartest model in that tier. It is that NVIDIA controls the hardware stack the model is tuned for, from consumer RTX cards through DGX systems to the H100/H200/B200 GPUs cloud providers rent by the hour, and it ships day-one optimizations (NVFP4 quantization, speculative decoding draft models) that a rival lab without NVIDIA’s hardware access cannot match as tightly.
That vertical integration is Lightning’s real differentiator. A developer choosing between similarly capable open models has a practical reason to pick the one NVIDIA itself tuned for NVIDIA GPUs, especially if that developer is already running on NVIDIA hardware for other workloads. It is a soft form of lock-in built on performance rather than licensing terms.
What Developers and Enterprises Should Watch Next
Teams evaluating Lightning for production agent pipelines should benchmark it directly against whatever model currently handles their execution-layer tasks, rather than trusting NVIDIA’s aggregate 4x speed claim at face value, since that figure is measured against unspecified “similar-sized models” rather than a named baseline. The PinchBench comparison against Qwen3.6-35B is more useful precisely because it names a specific competitor and a specific task count.
Cost modeling matters too. A free model still runs on hardware that costs money by the hour, so the real savings show up in throughput per dollar of GPU rental, not in the license fee NVIDIA is waiving. Our current H100, H200, and B200 cloud pricing breakdown is a useful starting point for teams trying to model what Lightning actually costs to run at scale on rented hardware.
Predictions: Where Nemotron Goes From Here
- Expect NVIDIA to keep splitting the Nemotron line into a fast execution tier and a separate reasoning-heavy planning tier, rather than merging the two into one general model.
- Expect more serverless inference providers to add Lightning to their menus within weeks, given how quickly the first wave (DeepInfra, Fireworks, FriendliAI, CoreWeave, and others) moved at launch.
- Expect rival chipmakers and cloud providers to respond with their own free execution-tier models tuned for their own hardware, following the same playbook NVIDIA just ran.
- Expect NVIDIA to fold Lightning-class models into more of its enterprise agent tooling, given the direct references to OpenClaw, Hermes Agent, and the NemoClaw management stack in the launch materials.
- Expect the next Nemotron release to arrive within NVIDIA’s now-established four-to-six-week cadence, likely pushing the Intelligence Index score higher again without a major jump in active parameter count.
Frequently Asked Questions
Is Nemotron 3.5 Lightning really free to use commercially?
Yes. It ships under the OpenMDW-1.1 license, which CNBC and Artificial Analysis both describe as permissive and open for commercial use without needing NVIDIA’s permission or payment.
What GPU do I need to run Nemotron 3.5 Lightning?
A single high-end consumer GPU such as an RTX 5090 can run it locally through tools like Ollama or LM Studio. For production or cloud deployment, NVIDIA and its partners list H100, H200, B200, and A100 shapes, plus DGX Spark and DGX Station for local data-center-grade inference.
How is Nemotron 3.5 Lightning different from Nemotron 3 Nano?
Lightning is the direct successor to Nemotron 3 Nano 30B A3B, at a similar parameter footprint but with a higher Artificial Analysis Intelligence Index score, 24 versus 15, thanks largely to its hybrid Mamba-2 and MoE architecture.
Is Nemotron 3.5 Lightning meant to replace a chatbot like ChatGPT?
No. NVIDIA positions it as an execution-layer model for always-on agents handling repetitive, specialized tasks, not as a general conversational assistant competing on broad reasoning ability.
How does it compare to Qwen3.6-35B?
On a 10,000-task PinchBench run, Lightning reportedly finished 30% faster than Qwen3.6-35B while holding accuracy around 86%, making it a direct speed-focused alternative in the same size class.
Where can I download Nemotron 3.5 Lightning?
Open weights are available on Hugging Face and ModelScope, with hosted access through OpenRouter, NVIDIA’s own NIM microservice on build.nvidia.com, Ollama, AWS SageMaker JumpStart, and Oracle OCI Generative AI.
Does Nemotron 3.5 Lightning support fine-tuning?
Yes. NVIDIA supports both LoRA and full supervised fine-tuning through its NeMo Automodel toolkit, plus reinforcement learning through NeMo RL and NeMo Gym, including a dedicated agentic training dataset.
Why would NVIDIA give away a model instead of charging for it?
Analysts covering the launch argue it is a demand-generation move: a free, fast model optimized for NVIDIA hardware encourages more always-on agent deployments, which in turn consume more GPU hours over time, supporting Nvidia’s core data-center hardware business even though the model itself carries no license fee.
