Nemotron 3.5 Lightning vs Nemotron 3: The Same-Line Upgrade Question

Nemotron 3.5 Lightning is a same-line successor to Nemotron 3, built around one fast 30B model instead of a size ladder. NVIDIA released it on August 11, 2026 as an open-weight model with 3B active parameters.

Nemotron 3, from December 2025, ships as a family: Nano at 3B active, Super at 12B active, and Ultra at 55B active. Both lines use a hybrid design and both reach up to 1 million tokens of context.

The short answer: upgrade for speed and cost on high-volume agent loops, stay for the size ladder. Both are open weights with no per-token fee when self-hosted, so a test is cheap to run.

This guide compares the two on architecture, speed, agentic focus, and NVIDIA's published benchmarks. It ends with a plain call on when to upgrade and when to stay put.

Nemotron 3.5 Lightning vs. Nemotron 3: Side-by-Side

Architecture: One Fast Model vs a Size Ladder

The core difference is shape: Nemotron 3.5 Lightning is a single 30B model, while Nemotron 3 is a family of three sizes. Lightning activates just 3B of its 30B parameters per token, an A3B Mixture-of-Experts pattern.

Both lines are hybrids, not plain Transformers. Lightning interleaves Mamba-2 state-space layers, MoE layers, and a few Attention layers. Nemotron 3 uses a hybrid Mamba-Transformer MoE across Nano, Super, and Ultra.

The design goals differ. Nemotron 3 lets you climb tiers when a task needs more capability. Lightning bets that one well-tuned small-active model can cover most agent work at higher speed.

Mamba-2 layers help here. They process long sequences with lower memory growth than pure Attention, which is part of why both lines reach 1M context. Lightning leans on that hybrid mix to stay fast at its small active size.

Context Window: A Tie at 1 Million Tokens

Both models handle up to 1 million tokens of context, so neither wins on raw window size. That is enough for large codebases, long agent histories, or many documents at once.

What differs is cost per long call. Lightning keeps only 3B parameters active, so long-context runs stay cheaper and faster on the same hardware. Nemotron 3 Super and Ultra activate more, which raises the price of very long sessions.

For agents that reread a growing memory on every step, active-parameter count matters more than the ceiling. Lightning's low active count is the practical edge here, not the window itself.

Do not treat 1M tokens as free. Filling the window still costs compute and time, so most agents work best with careful context management rather than dumping everything in. The high ceiling is a safety margin, not a default setting.

→ Compare per-token pricing for every AI model

Speed and Agentic Focus

Speed is where Lightning is meant to pull ahead. NVIDIA positions it for fast, accurate, specialized task execution inside long-running agents, and claims up to 4x output speed versus similar-sized models.

That framing is narrower than Nemotron 3's. The Nemotron 3 family is a general ladder you size to the task. Lightning is tuned for agent loops that run the same kind of step thousands of times.

In our work evaluating open-weight models for clients, the upgrade that pays off is the one tied to a real agent workload, not a benchmark score. A model that is 4x faster on a tight loop changes unit economics; a model that is 4x faster on a task you run twice a day does not.

Benchmarks: Read Them as NVIDIA's Published Numbers

NVIDIA publishes strong scores for the Lightning BF16 checkpoint. The headline figures are SWE-bench Verified 51.56, GPQA Diamond 75.44, MMLU Pro 81.94, and PinchBench 85.37.

Treat these as NVIDIA's own reported results, not independent tests. Nemotron 3 published its scores per tier at its December 2025 launch, and the tiers differ, so a fair read compares Lightning to a specific Nemotron 3 size.

The takeaway is to verify on the model card before you commit. Benchmarks move with checkpoint updates, and a 30B-A3B model beating or matching a larger tier on one test does not mean it wins on your task.

SWE-bench Verified is the most relevant score for coding agents, since it measures real software fixes. GPQA Diamond and MMLU Pro test reasoning and broad knowledge, while PinchBench adds another lens. Weigh the benchmark that maps to your actual workload, not the highest number.

License and Openness

Lightning ships under OpenMDW-1.1, and NVIDIA released the weights plus training data and recipes. That is a more open package than Nemotron 3, which uses the NVIDIA Open Model License.

For most business users, both are self-hostable open weights with no per-token fee. You pay for hardware and operations, not for tokens, unless you use a hosted endpoint.

The recipe and data release matters for teams that want to reproduce or extend the model. If you plan to fine-tune or audit training data, Lightning gives you more to work with.

Checkpoints, Training, and How to Run It

Nemotron 3.5 Lightning ships in two checkpoints: NVFP4, a 4-bit format, and BF16. The 4-bit build lowers memory needs, while BF16 keeps higher precision for quality-sensitive work.

NVIDIA pre-trained the model on more than 20 trillion tokens. You can run it as a hosted API on build.nvidia.com, locally on RTX or DGX hardware, or through open serving stacks like vLLM.

Nemotron 3 does not offer this exact checkpoint pair, so serving choices differ between the lines. If you want a 4-bit build to fit tighter GPUs, Lightning gives you that option out of the box.

NVIDIA also shipped NeMo Switchyard alongside the model. It is a separate model router, not part of the model, and it can direct requests across models in a mixed stack. That helps if you keep both Lightning and a Nemotron 3 tier in service and route by task.

When to Upgrade and When to Stay

Upgrade to Lightning when your bottleneck is agent speed or cost per step. A high-volume loop that runs the same task thousands of times is the clearest win, especially with the 4x speed claim and a 3B active count.

Stay on Nemotron 3 when you rely on the size ladder. If you route easy calls to Nano and hard calls to Ultra, one 30B model may not replace all three tiers cleanly.

Also weigh the switch cost. Both are open weights and both hit 1M context, so migration is mostly prompt tuning and serving changes, not a rewrite. Test Lightning on your real workload before you retire Nemotron 3.

A common middle path is to run both. Route high-volume, repetitive agent steps to Lightning for speed, and keep a Nemotron 3 tier for the rare task that needs more capability. NeMo Switchyard can handle that routing without a full re-architecture.

The Verdict

Pick Nemotron 3.5 Lightning if your main constraint is speed and cost inside long-running agents. Its 30B-A3B design and up to 4x output speed target exactly that job, and its more open OpenMDW-1.1 license and released recipes make it friendlier to fine-tune.

Stay on Nemotron 3, or keep it alongside, if you depend on the Nano, Super, and Ultra ladder to match model size to task difficulty. A single 30B model can simplify your stack, but it may not cover the top-end work that Ultra handles.

The safe path is a side-by-side test. Both reach 1M context and both are self-hostable open weights, so you can run Lightning against your current Nemotron 3 workload and compare speed, cost, and quality on tasks you actually ship. Verify all benchmark numbers on the official model card first.