Nemotron 3.5 Bolt 24GB vs Nemotron 3.5 Lightning 30B-A3B: Same Family, Opposite Deployments

Two models from one family, released four months apart, built to opposite ends of the same question. N​VIDIA N​emotron 3.5 Lightning 30B-A3B shipped in August 2026 as a high-throughput text model for data-centre serving. N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB shipped on 15 October 2026 as a multimodal agent that fits on a laptop. Bolt has double the total parameters and three times the active parameters of Lightning, and it runs on hardware Lightning cannot start on. This is what actually separates them, and where the family resemblance is misleading.

Both come out of the same training recipe

N​VIDIA's technical report for the pair is explicit that both models begin with the same two stages: a 100-billion-token warm-up for multi-token-prediction heads, then a trillion tokens of distillation from the next-larger N​emotron 3 model. Lightning distils from N​emotron 3 Super Base; Bolt distils from the N​emotron 3.5 Super VL line, which is why Bolt can see and Lightning cannot.

After that the paths separate, and the separation is the whole story.

Lightning goes through a 21-billion-token long-context phase that extends it to a million tokens, then an SFT-and-RLVR post-training pipeline, and it is released as a 30B total / 3B active text model with DSpark, DFlash and MTP speculative options and a single-GPU deployment target of one H100 or one DGX Spark with NVFP4 weights.

Bolt is compressed from a 120-billion-parameter parent by Star Elastic architecture search, re-attached to its vision tower, retrained through a vision-heavy supervised stage, then put through end-to-end reinforcement learning in agentic environments, and finally quantised to 2.0625 bits per expert weight with quantisation-aware distillation. It comes with one drafter, LiLiCorr, and one deployment shape: something with 24GB of memory and the ability to run it alongside other applications.

The spec contrast, side by side

• Parameters — Bolt: 74B total, 9B active. Lightning: 30B total, 3B active.

• Precision as shipped — Bolt: IQ2_XXS experts at 2.0625 bits per element, NVFP4 Mamba layers, Q8_0 embeddings and attention. Lightning: NVFP4 across the board for the deployment checkpoint, with a BF16 reference release for customisation.

• Context — Bolt: 262,144 tokens. Lightning: one million tokens, the result of a dedicated long-context training phase.

• Modality — Bolt: text, images and video in. Lightning: text only. The Artificial Analysis entry for each model says so on the same panel with the same wording, which makes it one of the few rows where the two are directly comparable without a caveat.

• Host hardware — Bolt: RTX Spark laptops, GeForce RTX PCs, DGX Spark, DGX Station, RTX PRO workstations. Lightning: a DGX Spark or an H100, or Blackwell and Ampere via W4A16.

• Speculative decoding — Bolt: LiLiCorr, a 1.13B drafter, seven drafted tokens per step. Lightning: DSpark for low-concurrency work, with DFlash and MTP also published.

• Licence — OpenMDW-1.1 for both, and both are cleared for commercial use.

The one place both are measured by someone else

These two models have something most of the N​emotron family lacks: entries on the same independent panel, measured the same way. That makes the comparison unusually clean for a family-internal matchup.

Artificial Analysis scores N​emotron 3.5 Lightning at 13 on its Intelligence Index, ranking 28th of 141 models, and N​emotron 3.5 Bolt 74B-A9B at 21, ranking fourth of 67 models in its own class. Those classes are not identical — Bolt is a medium open-weights model, Lightning is a small one — so the ranks are not directly comparable, but the index scores are produced by the same methodology on the same evaluations.

Speed is where the gap inverts, and this is the single most decision-relevant fact about the pair. The panel measures Lightning at 307 output tokens per second, first of 63 models on that dimension. It reports Bolt's speed as unavailable, because there is no public API serving it and the panel's speed measurement requires one. The only throughput figure that exists for Bolt is N​VIDIA's own: 83.99 tokens per second on AIPerf SPEED-Bench-Coding, on a Surface Laptop Ultra with 128GB of memory, with the drafter attached.

Read those together and the family story becomes concrete. Lightning is roughly three and a half times faster than Bolt on the numbers we have, in a form you can rent by the token; Bolt is meaningfully smarter on the same index and runs on a machine nobody can rent at all.

Three ways the choice resolves

If the workload is high-volume and text-only, Lightning is not a compromise, it is the better tool. A million-token context at 307 tokens per second with a published rate of six cents per million input tokens and twenty cents per million output is a specific and useful thing, and Bolt cannot match the speed or the context on hardware of any kind.

If the workload is visual, there is no decision to make. Bolt accepts images and video and Lightning does not, and no amount of context or throughput substitutes for a modality that is absent. The WebArena vision row in N​VIDIA's own table — 50.86 for Bolt, with Lightning not listed at all — is what that looks like when it moves from a spec sheet into a benchmark.

If the workload has to run without a network, Bolt is the only one of the two that has been packaged for it. Lightning's deployment target is a single H100 or DGX Spark with an NVFP4 checkpoint, which is a server; Bolt's is a 24GB file on a laptop that is also running a browser. The distinction is not about capability but about where the machine sits.

What the family gets out of shipping both

The interesting thing about this pair is not that one beats the other. It is that N​VIDIA built the cheap one first, then spent four months building the expensive one to fit somewhere the cheap one could not go.

Lightning is the throughput play: few active parameters, enormous context, designed to be served many times over on one GPU. Bolt is the density play: an enormous parameter count compressed into a fixed memory budget so that a single user on a single machine gets something closer to frontier agent behaviour locally. The two optimise different ratios — tokens per second per GPU against intelligence per gigabyte per device — and the fact that Bolt scores 21 against Lightning's 13 is not a defeat for Lightning, because Lightning would not fit on the machine Bolt was built for in the form Bolt exists in.

There is also a data point for the compression argument itself. Bolt's vendor table includes Lightning as one of its baselines, and Bolt beats it on nearly every agentic row: 50.98 against 23.46 on Terminal-Bench 2.1, 66.75 against 36.81 on BrowseComp, 17.01 against 9.48 on Tau3 Banking, 28.47 against 18.26 on GDPVal-AA-V2. Those are N​VIDIA's own numbers comparing two of its own models in the same table, and the honest reading is that compressing a 120B parent beats training a 30B model from scratch at the same target size. That is precisely the claim Star Elastic was invented to support.

Where this leaves anyone building an agent

The pairing suggests a shape rather than a winner. Lightning handles the high-volume, long-context, text-only middle of an agent's work where throughput and context are the binding constraints. Bolt handles the local, visual, latency-tolerant work that cannot leave the machine. Neither is a replacement for the other, and N​VIDIA's own framing of the family — Ultra for the hardest reasoning, Lightning for throughput, Bolt for the edge — is an admission that the boundary is real.

What neither model solves is the part of the stack that is not local and not N​VIDIA's. We do not host either checkpoint; the OrcaRouter catalogue has no N​emotron entry, so there is no endpoint or rate to quote for these two. What a gateway contributes to a mixed stack is the other side: one key across 200-plus hosted models, provider list price passed through with no markup added, automatic failover when a provider degrades, and a routing DSL for describing escalation across models as configuration. The choice between Bolt and Lightning is made on your hardware and in your harness. What happens when the task exceeds both, or when a request has to reach a model neither of them can be, is the part that lives behind an endpoint.

Which one to try first

If you have a GPU with 24GB and a network connection that comes and goes, Bolt is the more interesting download: it is newer, it is multimodal, and it does something no other release in the family can.

If you have a DGX Spark, an H100, or an application where latency is the product, Lightning has been generally available since August, has an independent speed measurement that puts it first in its class, has a published price, and has had two months for the ecosystem to find its sharp edges. Bolt has had one day, its drafter is a mandatory third download, and nobody outside N​VIDIA has measured its latency.

Both statements can be true at once. That is what a family is for.

That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.