Nemotron 3.5 Bolt 74B vs Nemotron 3 Ultra: 1.5 Index Points, Seven Times the Parameters

The most surprising number in the N​VIDIA N​emotron 3.5 family is not the size of its flagship. It is that N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB, a 74-billion-parameter checkpoint squeezed into 24 gigabytes at two bits per weight, lands within 1.5 points of N​emotron 3 Ultra, the 550-billion-parameter model that the rest of the family was distilled from, on Artificial Analysis's independently run Intelligence Index — 21.4 against 22.9. Seven times the parameters, six times the active compute per token, and a gap you could lose in a benchmark's error bars. Bolt's weights appeared on Hugging Face on 7 October 2026 with N​VIDIA's announcement following on 15 October; Ultra shipped on 4 June 2026.

The two models are not interchangeable, and the places where they differ are more instructive than the headline score. What the compression preserved and what it discarded are both visible in the published numbers, and they are not the same rows.

What each one is

Ultra is the flagship of the previous generation and the teacher at the centre of N​VIDIA's distillation pipeline. It is a 550B-total, 55B-active Mamba-2 hybrid with Multi-Token Prediction, released under OpenMDW-1.1 with its training data and recipes, built for long-running agentic work with up to a million tokens of context and a 65,000-token output cap per request. N​VIDIA's own claim for it was the highest-intelligence US open-weights model of its release window, and the independent record largely supports that: the panel scores it 22.9 in its reasoning configuration, with a measured 156 output tokens per second, 1.29 seconds to first token, and a median 19 seconds of end-to-end response time.

Bolt is the edge tier of the newer generation, and it is not a distilled-Ultra in the direct sense. Its parent is N​emotron 3.5 Super VL 120B-A12B, the 120-billion-parameter multimodal model N​VIDIA launched alongside it. The compression was done by Star Elastic, N​VIDIA's router-guided architecture search, which optimised for the lowest knowledge-distillation loss that satisfied a hard footprint of under 24GB at batch size 1 and 128K context — pulling expert top-K from 22 down to 9 and quantising expert weights to IQ2-XXS at 2.0625 bits per element, with quantization-aware distillation to recover accuracy. It carries a C-RADIO v4-H vision encoder because its parent was multimodal.

The independent rows, both sides

Everything below is from Artificial Analysis unless marked otherwise, which is what makes the comparison usable: both models were run by the same evaluator under the same harness.

• Intelligence Index — Bolt 21.4; Ultra 22.9. Independently measured for both.

• GDPval-AA (normalised) — Bolt 0.240; Ultra 0.257. A 1.7-point gap on a benchmark built from real occupational deliverables.

• SciCode — Bolt 0.397; Ultra 0.403. Effectively level.

• Humanity's Last Exam, no tools — Bolt 24.7 percent; Ultra 28.4 percent. The four-point gap here is the largest on this list in relative terms.

• AA-LCR long-context retrieval — Bolt 69.7 percent; Ultra 79.3 percent. Ten points, and the clearest evidence of what 2-bit compression costs.

• AA-Omniscience — Bolt 17.1 percent accuracy with 56.3 percent non-hallucination; Ultra 22.6 percent accuracy with 70.3 percent non-hallucination. The flagship is both more knowledgeable and better calibrated, which is the least surprising result on the page.

• Terminal-Bench 4.0 (harder set) — both score zero. Neither model is solving that tier, flagship included.

• Context window — Bolt: up to 256,000 tokens per the model card, and 262,144 on the panel's record. Ultra: up to 1,000,000 per N​VIDIA, with a 65,000-token maximum output.

• Inputs — Bolt: text, images and video in, text out. Ultra: text in, text out. Nothing in the flagship's release materials claims a vision tower.

• Price — Bolt: none. There is no published metered rate for the 24GB checkpoint, and the panel lists no input or output figure for it. Ultra: $0.60 per million input tokens and $2.50 per million output, with cached input at $0.16, and a measured $0.60 of spend per Intelligence Index task.

• Speed — Bolt: no independent measurement. Ultra: 156.1 output tokens per second, 1.29 seconds to first token, median 14.6 seconds of reasoning time per response.

• Footprint — Bolt: 24GB at IQ2, deployable on one A100, H100 or Blackwell part. Ultra: a multi-accelerator deployment; N​VIDIA's own recipe for the teacher checkpoints in this family names four B200s or eight H100s as a floor, and Ultra is the largest member.

Why the index gap is smaller than the size gap, and what that means

There are two honest explanations and both are probably true in part.

The first is that the Intelligence Index is a composite that rewards procedure more than recall. Bolt lost its stored facts to 2-bit quantisation — you can see it in the accuracy rows — but kept enough of the chain-of-thought, tool-calling and structured-output behaviour to score on the composite tasks. Note that Ultra's advantage concentrates exactly where memory matters: long-context retrieval, omniscience accuracy, hallucination rate, hard science questions that require a fact rather than a method.

The second is that architecture search on a strong parent is a better use of a fixed budget than training small from scratch, which is N​VIDIA's explicit argument in the Bolt documentation. Taking a 120B multimodal teacher and searching for the widest student that fits 24GB is a different exercise from pretraining a 30B model, and the resulting 74B/9B checkpoint sits where an unusually efficient 74B would.

What that does not mean is that Bolt replaces Ultra. A model that admits uncertainty more often (56.3 versus 70.3 percent non-hallucination, in the wrong direction) and retrieves from long context less reliably is a worse assistant for the questions people actually ask. It is a better assistant for tasks that are mostly procedure: calling tools, filling forms, editing files, reading a screenshot and clicking the right thing.

The two things Ultra cannot do

Bolt's advantages are narrow, and both are structural rather than numerical.

It fits. A 24GB checkpoint runs on hardware a person owns, with no network, no per-token meter and no capacity negotiation. Ultra's intelligence is only reachable over an API, at $0.60 in and $2.50 out per million tokens and roughly $0.60 of spend per index task. That price is genuinely low for a model of its class, which is why the hosted route is a real option rather than a compromise — but it is still a route that requires connectivity and trust in someone else's uptime.

It sees. Bolt's C-RADIO v4-H vision encoder gives it image and video input, and N​VIDIA's model card reports 50.86 on WebArena vision and 66.75 on BrowseComp. Ultra is a text model. If the agent's job involves screenshots, PDFs, dashboard reading or browser automation with visual state, the flagship of the older generation cannot be substituted in at any price.

How to choose

Take Ultra if the workload is reasoning-heavy and text-only: multi-step analysis, long document sets, coding over large repositories, or anything where being wrong is expensive. Its independent scores beat Bolt nearly everywhere, the price is among the lowest in the high-intelligence bracket, and its throughput is measured rather than assumed.

Take Bolt if the workload runs on hardware you control, needs to see, or has to keep running when the network does not. The 1.5-point index gap is real and narrow enough that the decision is usually made on deployment rather than on quality — and the fact that N​VIDIA shipped a 24GB file that scores within two points of its own 550B flagship is the more newsworthy fact of the two releases.

There is a middle path worth naming, because it is how N​VIDIA positions the family rather than how a single-model buyer would: a small local executor doing tool calls and screen reading, with a larger hosted model handling planning and hard reasoning. That architecture needs a reliable way to reach the hosted half across more than one provider, which is what OrcaRouter's failover is for — automatic retry across providers behind one endpoint, on a key that also reaches 200-plus models at the provider's list price with 0 percent markup added. We do not host any N​VIDIA model; the N​emotron family is not in our catalogue, and nothing here is a claim to serve it. The point is only that a two-tier agent design needs the second tier to be routable, and a single hardcoded vendor endpoint is the usual way that design fails.

What would change this page

Two measurements. The first is an independent throughput and latency run on the 24GB checkpoint, which would either confirm or destroy the case for using it as a local executor at interactive speed; today the only number available is N​VIDIA's own 83.99 tokens per second on AIPerf SPEED-Bench-Coding, measured on a laptop-class machine and not comparable to the panel's rig. The second is an independent re-run of the agentic benchmarks on the 2-bit weights at a quantisation that a third party controls — the vendor table has Bolt beating Ultra on Terminal-Bench 2.1 by nearly 27 points, and that claim is doing an enormous amount of work without outside confirmation.

None of this is a reason to avoid the model, only a reason to know the history: the documented AI incidents and agent failures behind 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.