Nemotron 3.5 Bolt vs Nemotron 3.5 Lightning: What NVIDIA's Own Table Says When You Read Both Columns

NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB and NVIDIA Nemotron 3.5 Lightning 30B-A3B are two tiers of the same family, released ten weeks apart, and the vendor has already published a table putting them head to head. That is unusual and useful: most head-to-head tables you read are assembled by a journalist or a vendor picking flattering comparisons, and this one is a lab benchmarking its own newer model against its own older one, in public, with the older one named as a baseline. It is still vendor-reported and nobody outside NVIDIA has reproduced a row of it — but the intent behind it is not in doubt, and neither is the direction of the result. Bolt takes twelve of the thirteen rows. It is worth understanding exactly which one it loses and why that row is the interesting one.

The family structure matters before the numbers. Nemotron 3.5 Lightning shipped on 11 August 2026 as the execution layer of an agent stack — a 30B-total, 3B-active hybrid that NVIDIA distilled from the far larger Nemotron 3 Ultra specifically so a frontier model could do the planning while Lightning grinds through tool calls on a single GPU. Bolt is a different lineage entirely: it was compressed from the 120B-A12B Nemotron 3.5 Super VL model, not from Ultra, and it exists to be a vision-language generalist that fits in 24GB. Both are open-weight under OpenMDW-1.1, both are sparse mixtures of experts on Mamba-2 hybrid backbones, and both ship speculative decoding — Lightning with Multi-Token Prediction heads plus DFlash and DSpark drafters, Bolt with a LiLiCorr drafter driven through DFlash that proposes seven tokens ahead.

The shared lineage stops being useful at the point you have to decide, because the two models answer different questions. One is a cheap executor. The other is a local generalist.

Reading the table by role, not by score

Every figure below comes from NVIDIA's own model card for the Bolt 24GB checkpoint, which lists Lightning as a baseline. The Bolt column is IQ2, a 2-bit quantised artifact. The Lightning column is BF16, unquantised. That asymmetry is the single most important thing to know about the table, and it cuts in Bolt's favour — it is winning those rows while compressed roughly eight-fold relative to its parent.

• General agentic reasoning — GDPVal-AA-V2: Bolt 28.47, Lightning 18.26. A ten-point gap, and the row closest to "how good is this at the everyday office-shaped work the model was built for."

• Terminal work — Terminal-Bench 2.1: Bolt 50.98, Lightning 23.46. This is the widest gap on the board and it is not a surprise, because NVIDIA's own guidance for Lightning is that it should not be the primary agent. An independent evaluation recorded roughly 24.6 for Lightning on the same benchmark, which is close enough to the vendor's 23.46 to suggest the number is real.

• Scientific and knowledge reasoning — GPQA Diamond: Bolt 86.43, Lightning 75.57. SciCode: Bolt 41.02, Lightning 31.38. Humanity's Last Exam without tools: Bolt 25.30, Lightning 10.47. Consistency across three different shapes of hard question.

• Long-context reasoning — AA-LCR: Bolt 64.06, Lightning 49.19. Fifteen points, and here the memory constraint is not the limiting factor for either model.

• Web and browsing agency — BrowseComp: Bolt 66.75, Lightning 36.81. Thirty points. The largest single gap in the table, and the one that most directly contradicts the idea of Lightning as an autonomous executor.

• Tool-use density — MCP-Atlas: Bolt 52.20, Lightning 46.4. PinchBench: Bolt 87.82, Lightning 83.43. Both closer than the gaps above, which is consistent with Lightning's actual design brief: short, high-frequency tool steps.

• Vision — WebArena vision: Bolt 50.86, Lightning has no entry. This is not a reporting gap. Lightning is text in, text out; it has no vision encoder to score.

The one row Lightning wins, and why it is not noise

AA-Omniscience non-hallucination: Nemotron 3.5 Lightning 68.81, Nemotron 3.5 Bolt 61.27. On the accuracy half of the same evaluation the two are within a third of a point of each other — 16.90 for Bolt against 16.63 for Lightning — which makes the gap on the non-hallucination half the whole story rather than a rounding artifact. The smaller, older, unquantised model is measurably more restrained about saying things it does not know.

That is a real finding and it has a plausible mechanism. Bolt reached its footprint through aggressive 2-bit quantization with quantization-aware distillation; the lossy part of that pipeline is exactly the part that governs calibrated uncertainty. Lightning is smaller but sits at full BF16 precision, and its distillation target was a reasoning model rather than a compression ratio. If your agent's failure mode is a confident wrong answer that propagates into an action, this is the row to weigh heaviest — and no amount of routing cleverness fixes a model that invents more.

Vendors rarely publish a table containing a row their new model loses. That this one does is a small piece of evidence that the table is not purely promotional.

Serving cost is the real second axis

The scores above are not the whole decision, because these two models are not equally expensive to run.

• Active parameters — Bolt activates 9B per token; Lightning activates roughly 3B. Three times the compute per token for the larger model, on every token, forever.

• Memory envelope — Bolt is a 24GB file with everything already compressed; Lightning is a 30B BF16 checkpoint that wants a single GPU and fits comfortably on an RTX 5090, a DGX Spark or a Jetson part.

• Context — both are specified to 1M tokens, and both are practically narrower. Lightning serves 256K on a single H100 before the memory maths gets awkward; Bolt's published window is 262,144, and that window shares the same 24GB with the KV cache, the images and the tool results.

• Speed claims — NVIDIA claims up to 4x output speed versus similar-size models for Lightning through its MTP and DFlash/DSpark drafters, and at least 80 tokens per second of decoding as the design target for Bolt, with 84 measured on the AIPerf SPEED-Bench on RTX Spark and 83.99 in a llama.cpp run on a Surface Laptop Ultra with 128GB. Both are vendor figures. Neither has been reproduced by anybody outside NVIDIA.

• Metering — neither has a hosted price from us or from any platform we can point you at. The cost is your hardware either way, which is precisely why the active-parameter count matters more than it does in a hosted comparison.

What the panel could and could not measure

Artificial Analysis has an entry for Nemotron 3.5 Bolt 74B-A9B and scores it 21 on the Intelligence Index, fourth of 67 in its footprint class against a class median of 8. Its speed and cost-per-task fields read N/A, because there is no hosted endpoint to call. Those two blanks are the honest state of the model: its intelligence floor is independently verified and its performance ceiling is not.

There is no Artificial Analysis entry for Nemotron 3.5 Lightning at the time of writing, which leaves the whole comparison resting on NVIDIA's own card. The independent datapoint that does exist — an LLM Stats verification of roughly 24.6 on Terminal-Bench 2.1, consistent with the vendor's 23.46 — supports the vendor's numbers rather than contradicting them. That is reassuring about the table without being a substitute for a second harness.

How to actually choose

The design intent decides most cases before the benchmarks do. Lightning is an executor by construction: NVIDIA's own guidance is to let a frontier model plan and Lightning run the loop, and its best rows in the table are the high-frequency tool-call ones rather than the autonomous-agent ones. Bolt is a generalist by construction: the same table shows it far ahead on browsing agency and terminal work, which are the planner-ish tasks Lightning was never meant to own — and it is the only one of the two that can see.

The composition that both models were built for is a planner, an executor and a local specialist. Because neither checkpoint carries a hosted price, the split lands on your own hardware and one endpoint in front of it does the useful work: a routing rule that sends short tool-call loops to a Lightning deployment, the vision-bearing and browser-driving requests to Bolt, and anything that needs genuine frontier reasoning out to a hosted model — all behind one key, with failover across upstream providers when a hosted leg degrades. The rule lives in configuration rather than in every service, which is the difference between a routing layer and a rewrite.

The open question is a single independent run. If a third party reproduces the Terminal-Bench and BrowseComp gaps on Bolt, the 24GB checkpoint is genuinely the stronger generalist and the case for the 30B tier narrows to pure throughput economics. If it does not, the honest reading is that a 2-bit quantised 74B and a BF16 30B are closer than one vendor's table implies — and that the non-hallucination row, the one Lightning wins, deserves more weight than its position at the bottom of the card suggests.