Nemotron 3.5 Bolt vs Nemotron 3 Ultra: What Survives a 7.4-Times Compression

Two models from the same vendor, separated by a factor of 7.4 in parameter count and by roughly two orders of magnitude in the hardware needed to run them. NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB was released on 15 October 2026 as a 24GB file that runs beside your other applications on a Spark device or a Blackwell GPU. NVIDIA Nemotron 3 Ultra 550B-A55B has been on Hugging Face since 3 June 2026 and asks for sixteen H100s at the floor in BF16.

The interesting question is not which is stronger — the answer is obvious and not close. It is what specifically survives when you compress a large model into a small memory envelope, because that tells you which workloads the small one can absorb and which ones will embarrass it. NVIDIA's own research note answers part of that question directly, and the parts it does not answer are where the real decision lives.

They are not parent and child, whatever the naming suggests

The obvious assumption is that Bolt is a shrunken Ultra. It is not, and getting this right changes what you expect from the file. Nemotron 3.5 Lightning, the 30B-A3B tier from 11 August 2026, was distilled from Ultra. Bolt was not. Bolt is compressed from Nemotron 3.5 Super VL, a 120B-A12B vision-language model, using a router-guided search NVIDIA calls Star Elastic, followed by quantization-aware distillation down to 2.0625 bits per expert weight in an IQ2_XXS format.

That lineage explains the one difference that matters more than any benchmark row: Bolt has a C-RADIO v4-H vision encoder and accepts images and video. Ultra is text in, text out, and it is not a reporting gap — the checkpoint has no vision tower to score. The smaller, newer, heavily quantised model sees; the largest reasoning model in the family does not.

What the compression kept

Bolt carries 74 billion parameters with 9 billion active per token on a Mamba-2 and Transformer hybrid backbone. Ultra carries 550 billion with 55 billion active. The active-parameter ratio is 6.1 to 1, and the total-parameter ratio is 7.4 to 1, which means the compression removed proportionally more total capacity than per-token compute — the model got narrower faster than it got shallower.

What that produces, on the evidence available:

• Independent intelligence — Artificial Analysis scores Nemotron 3.5 Bolt at 21 on its Intelligence Index, fourth of 67 in its footprint-filtered class against a class median of 8. Ultra is not in the same class and scores far higher on the same index version. This is the compression's measurable cost.

• Independent speed — Artificial Analysis measures Ultra at 156.1 output tokens per second, tenth of 74 models on that page and in its top speed band. For Bolt the same fields read N/A, because there is no hosted endpoint to call. NVIDIA's own figures — 84 tokens per second on AIPerf SPEED-Bench on RTX Spark, 83.99 in llama.cpp on a Surface Laptop Ultra — are vendor-reported and unreproduced.

• Instruction following — Ultra records 82.9 on IFBench. This is the row to watch in an agent that has to obey a long tool policy, and it is the kind of capability that quantization-aware distillation preserves unevenly.

• Long-context reasoning — Ultra records 75.2 on AA-LCR against a 1M-token specification. Bolt's published window is 262,144 tokens, sharing 24GB with the KV cache, the images and the conversation. A wide declared window and the ability to use it are separate properties; here the larger model has more of both.

• Factual restraint — Ultra records 22.1 on AA-Omniscience hallucination with 77.91 non-hallucination. Bolt's own card records 61.27 on non-hallucination, the single row where it loses to the far smaller Nemotron 3.5 Lightning. Two independent signs pointing the same way: this is where aggressive 2-bit compression costs the most, and no routing trick repairs a model that invents more.

The deployment envelope, which is where the two actually separate

This is the part of the comparison that does not reduce to a score.

• Memory — Bolt is a single 24GB artifact, already quantised, meant to coexist with a browser and an IDE on one machine. Ultra in BF16 asks for sixteen H100s, eight H200s or eight GB200-class nodes per its own model card; the NVFP4 build narrows that to eight H100s at the floor. The difference between those two sentences is the difference between a purchase order and a procurement programme.

• Context configuration — Bolt serves 262,144 tokens directly in its published vLLM recipe. Ultra's recipes default to 256K and require an environment flag to reach the full 1,048,576, which is why the panel records 262K for Ultra despite a 1M specification. Read the specification, then read the recipe.

• Serving stack — Bolt runs under vLLM on Ampere, Hopper and Blackwell, or under llama.cpp on RTX Spark and DGX Spark hardware, under the OpenMDW-1.1 licence that permits commercial use. Ultra runs under the same licence on datacentre-class parallelism. Both are open weights; only one of them is portable.

• Prompt cost — Ultra is metered at $0.60 per million input tokens and $2.50 per million output tokens with a 73% cache discount, which comes to about $0.60 to run one Intelligence Index task — genuinely cheap for a model of that size, and available because someone else operates the cluster. Bolt has no price at all: no hosted endpoint, no vendor API, no rate card, and N/A in both the speed and cost fields of the only independent panel that lists it.

Reading the announcement for what it does and does not say

Bolt's checkpoint was on Hugging Face from 7 October 2026 with only a model card attached. The vendor announcement that names it — alongside a 4GB sibling, Nemotron 3.5 Bolt E4B, aimed at game characters and dated for 23 October — came on 15 October. Landing the weights eight days before the prose is not unusual, but it does mean a week of the public record contains no explanation of the model at all, and that any account written in that window was working from the card alone.

The announcement is also the place to look for what is absent. There is no hosted endpoint, no API, no rate card, no third-party platform metering Bolt — and no independent benchmark run on the 24GB weights by anyone. The vendor's numbers are real and specific, taken with llama.cpp on a Surface Laptop Ultra with 128GB; they simply have not been reproduced outside NVIDIA.

Where each model belongs, and where a router belongs

Neither of these models is on OrcaRouter's catalogue. We do not host any NVIDIA checkpoint, there is no Bolt endpoint and no Ultra endpoint, and nothing here is an offer to serve either one. What a single endpoint does buy, if you are running one or both yourself, is a way to keep the local and hosted halves of a workflow under one key — the 200-plus models on our catalogue sit behind one API key and one bill, at each provider's list price passed through with 0% markup added, so a vendor price change is live on our side the same day rather than at the next renewal. A request that needs a full million-token window or a frontend-scale reasoner goes out to a hosted model; the local file keeps the high-volume, privacy-sensitive work that should never have left the building. Automatic failover means a degraded upstream moves the request instead of failing it.

The placement follows from the deployment maths rather than from the scores. Bolt belongs on a workstation when the data cannot travel, when the workload is video-bearing, or when volume makes per-token billing the dominant line item — and when a 262K window and an already-visible factual-restraint gap are acceptable. Ultra belongs on a metered endpoint when the work needs a million-token reach, the strongest instruction following in the family, and the lowest hallucination rate NVIDIA currently publishes.

Who should act, and who should wait

Act now if the workload is local and video-bearing: Bolt is the only model in this family that both fits 24GB and can see, and no other open checkpoint at that footprint scores 21 on the independent index. Act now if the workload is long-context text reasoning at volume: Ultra is metered, measured, and cheaper per Intelligence Index task than a model of its size has any right to be.

Wait if your decision hinges on throughput. Bolt's speed is the one dimension with no independent measurement anywhere, and for a locally deployed model the speed is what you feel every second of every day. A single third-party latency run on the 24GB weights would settle more than the next ten vendor benchmark rows — and until it exists, capacity planning for Bolt is planning around a claim.