On August 11, 2026, NVIDIA released the open-weight large language model (LLM) "Nemotron 3.5 Lightning." The total parameter count is 30B (30 billion), but the parameters activated for processing each token are limited to 3B (3 billion). The goal isn't to take first place on the hardest reasoning benchmarks. Rather, it's to serve as the executor that quickly handles the massive volume of repetitive tasks—tool calls, result verification—within long-running AI agents.
To support this division of labor, NVIDIA also released the model router "NeMo Switchyard." Complex planning is sent to a large model, while highly routine execution is delegated down to Lightning. Rather than judging model superiority by a single overall score, this is a design that selects the necessary capability and processing time for each job.
Only 3B Out of 30B Runs—Trimming Compute at the Same Size
Nemotron 3.5 Lightning is a hybrid model combining Mamba-2, Mixture-of-Experts (MoE), and Attention. According to the config file, there are 128 routed experts plus 1 shared expert, with 6 selected per token. While the model as a whole retains substantial knowledge capacity, it doesn't compute all weights every time, making it lighter to run than a dense 30B model.
The previous-generation Nemotron 3 Nano from the same lineage was also a hybrid MoE with 30B total parameters and 3.5B active. This time, the change is less about shrinking size and more about refining the training and inference pathways for execution-oriented use cases. In NVIDIA's unified evaluation, scores rose from the previous generation: SWE-bench Verified went from 34.08 to 51.56, Terminal-Bench 2.1 from 8.29 to 24.58, and PinchBench from 66.11 to 85.37.
Distribution formats are also split by role. The BF16 version is a reference weight for customization such as supervised fine-tuning (SFT), reinforcement learning, and distillation, while the NVFP4 version is intended for inference deployment. Instructions are provided for running the NVFP4 version on a single DGX Spark or a single H100. According to NVIDIA's support matrix, Blackwell and Hopper support NVFP4, while Ampere supports W4A16. The license is OpenMDW-1.1, which permits commercial use.
The maximum context length is 1 million tokens. However, this doesn't mean the same headroom is available in every configuration. The BF16 version's single-H100 example is capped at 256K, while the 1-million-token example is shown for 8x H100 or GB200 setups. In actual operation, allocation of concurrency and cache capacity is what matters most first.
Time to Finish the Job, Ahead of Raw Performance
Lightning isn't a model that consistently outperforms the similarly-sized Qwen3.6 35B A3B in capability benchmarks. Even in NVIDIA's own unified evaluation, Qwen leads in many areas including knowledge, hard reasoning, and coding agent tasks. On the other hand, Lightning outperformed on IFBench, which measures instruction following.
| Benchmark | Nemotron 3.5 Lightning | Qwen3.6 35B A3B |
|---|---|---|
| MMLU Pro | 81.94 | 85.63 |
| GPQA Diamond (no tools) | 75.44 | 83.40 |
| SWE-bench Verified | 51.56 | 70.12 |
| Terminal-Bench 2.1 | 24.58 | 44.38 |
| PinchBench | 85.37 | 88.07 |
| IFBench (loose) | 71.88 | 63.71 |
Given this gap, it's hard to view Lightning as a standalone commander capable of handling complex planning on its own. In the independent Artificial Analysis Intelligence Index, its listed score is 24, falling short of Qwen3.6 35B A3B's 32. It's a model meant to be assigned tasks where the gap with Qwen falls within an acceptable range, shortening the time to completion.
Regarding speed, NVIDIA claims output speeds up to 4x faster than similarly-sized models. On PinchBench, it maintained approximately 86% accuracy while completing 10,000 tasks 30% faster than the similarly accurate Qwen3.6 35B. However, the "up to 4x" figure is NVIDIA's own claim, and Artificial Analysis's current model page does not list output speed. When deploying, completion times need to be re-measured using one's own prompt lengths and concurrency levels.
The Large Model Plans, Lightning Executes
Long-running agents don't just solve a hard problem once and finish. They repeatedly call tools, inspect return values, hand work off to sub-agents, and format outputs. If everything is routed to an expensive large model, the small delay and compute cost of each individual call accumulate across the entire workflow.
NeMo Switchyard is an open-source library for selecting which model to use for each request. In the configuration NVIDIA envisions, planning and complex reasoning go up to a frontier model, while execution goes down to Lightning. Lightning's value isn't determined by taking first place overall, but by whether it can shorten the completion time of the entire agent workflow while reducing the number of calls to the large model.
There are also acceleration mechanisms built into the model itself. Multi-Token Prediction (MTP) is trained to predict multiple future tokens rather than just the next one. NVIDIA additionally distributes draft models called DSpark and DFlash, used for speculative decoding where candidate tokens are generated in batches and verified by the main model. DSpark targets DGX Spark and low-concurrency data center use cases, while MTP mainly targets medium-to-high concurrency.
NVFP4 quantization also widens the range of speed and deployable hardware. In published evaluations, compared to the BF16 version, the NVFP4 version's PinchBench score dropped from 85.37 to 83.43, and Terminal-Bench 2.1 dropped from 24.58 to 23.46. Meanwhile, SWE-bench Verified rose from 51.56 to 52.80, and GPQA Diamond rose from 75.44 to 75.57—the change after quantization varies by benchmark. This is precisely why accuracy degradation shouldn't be summarized with a single average figure, but instead checked against the specific tasks being used.
How Far Can Under-100-USD Tuning Go?
CodeRabbit tested Lightning on the narrow but high-volume task of code review triage. From data derived from public repositories, they created a training set of 9,996 examples, and used 1,000 examples separated by repository as a fixed evaluation set. The exact-match rate against the reference configuration rose from the conventional GPT-class model's 75.8% to 80.4% after SFT, and reached 80.7% with a combination of SFT and RLVR (reinforcement learning with verifiable rewards).
The experiment took under 3 hours and cost less than 100 USD. On a single A100 80GiB processing 8 concurrent requests, total output reached 314.82 tokens per second. The estimated cost to process the same 1,000 examples dropped 50.4%, from 2.34 USD to 1.16 USD. For code review triage, tuning under 100 USD exceeded the match rate of the conventional model while nearly halving inference costs.
Still, the scope of applicability is limited. The 0.3-point increase in match rate from adding RLVR isn't statistically conclusive, and CodeRabbit itself has not verified sustained load or production traffic. The approach is to first carve out a repetitive task and confirm via fixed evaluation whether accuracy close to that of a large model can be achieved. Only afterward should completion time under concurrency, empty responses, and retry behavior on failure be measured. For teams that can follow this sequence, Lightning looks poised to become not a substitute for large models, but an execution engine for running agents longer and more cheaply.