Nemotron 3.5 Bolt 74B Release Date: NVIDIA Squeezes a 74B Agent Model Into 24GB

NVIDIA Nemotron 3.5 Bolt 74B-A9B ships on Thursday, October 15, 2026, and NVIDIA Nemotron 3.5 Bolt E4B follows on Friday, October 23. Those two dates are the answer to a tease: on October 14, NVIDIA's AI account posted "So many models launching this week! And we have one of our own coming...⌛️" with no name attached. Checking that tease is worth the two minutes, because it resolves in an unusual way — the announcement the tweet gestures at had already been published, in full, the day before, in an NVIDIA blog post dated October 13 that names both models, both dates, the memory footprint and the inference runtimes. So this is not a leak, and it is not a rumour. It is an under-promoted launch, and the tease was the nudge, not the scoop.

What follows separates what NVIDIA has actually written down from what nobody outside the company has measured yet — which, for a model that ships today, is most of the interesting column.

The tease, and what it resolved to

Two things are worth pinning before anything else, because they are exactly the kind of detail that gets flattened in aggregation. First, the ordering. NVIDIA's blog post "All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark" is timestamped 2026-10-13T12:00Z. The "one of our own coming" post came in at 2026-10-14T20:29Z — roughly thirty-two hours later. The tweet is a pointer, not the primary source, and any write-up that treats the tweet as the announcement is working from the wrong document.

Second, the phrase "one of our own" is doing real work. NVIDIA is not only the vendor here; it is the platform vendor for the week. The same October 13 post ships RTX Spark laptops on October 16, a rebuilt DLSS 5 AI model debuting in NBA 2K27 the same day, NVIDIA PAIR out of beta, and Broadcast 2.3. Bolt 74B is the language model in a week where NVIDIA is announcing several models of very different kinds. If you came looking for one product, that is why the coverage is confusing.

What NVIDIA has committed to, in its own words

The core claim is short and specific. NVIDIA describes Nemotron 3.5 Bolt 74B-A9B as "a new LLM purpose-built for running local agents on RTX Spark," built as "a 74B parameter, 9B-active-parameter mixture-of-experts (MoE) model" that is "packed into 24GB using 2-bit quantization with advanced quantization-aware training techniques to retain quality."

Three things in that sentence are doing the work:

• 9B active, not 74B. This is a sparse MoE. Routing cost per token tracks the active path, not the total parameter count, which is why a 74B model is being pitched for a laptop at all.

• 2-bit, deliberately. Not a community quantisation of a BF16 release, but a model NVIDIA says was trained with quantisation-aware techniques aimed at holding quality at that width. Whether it holds is a measurement, not a claim — see the ledger below.

• 24GB is the target envelope. NVIDIA frames the trade as "the optimal balance between speed and intelligence, being faster than most popular dense models while smarter than most popular MoE models with a similar memory footprint." That is NVIDIA's own positioning, unreproduced by anyone.

On access, NVIDIA names the runtimes explicitly: llama.cpp, vLLM, Ollama, Unsloth and LM Studio at launch, plus direct use through agents including Hermes and OpenClaw. And on licensing, the post says Bolt 74B is "fully open source... including training data, weights and recipes" — the same framing NVIDIA applies to the rest of the Nemotron line.

Why 24GB is the number that matters, not 74B

The headline integer here is 74B. The load-bearing one is 24GB, and for a straightforward reason: 24GB is a desktop-class accelerator envelope. It is the capacity of a single high-end consumer card, and on the laptop side it is the frame NVIDIA is building the week around — RTX Spark machines cap at up to 128GB of unified memory and 1 petaflop of AI compute, and the laptops themselves arrive October 16.

Put those together and the pitch is legible. A 9B-active MoE whose weights land in 24GB at 2-bit means the same build that fits a workstation card also fits the laptop shipping the next day, with headroom left over for the application and a second small model. That is a materially different deployment story from a hosted frontier model behind an API, and it is the reason "local agent" is in NVIDIA's own one-line description rather than "cheap inference."

The counterweight is that a 2-bit target is a constraint on what the model can be. Any quality comparison against a BF16 model of similar total size is a comparison across different artefacts, and NVIDIA chose the artefact on purpose. Judge Bolt 74B against other 24GB-class local builds, not against a full-precision checkpoint that would never have fitted.

The ledger: confirmed, stated-but-unmeasured, and unknown

Being precise about which column a fact sits in is the whole value of reading this in the first week rather than the third.

• Confirmed by NVIDIA: the model name, the 74B total / 9B active MoE shape, the 24GB 2-bit footprint, the October 15 date, the five named runtimes, agent access via Hermes and OpenClaw, and the assertion that weights, training data and recipes are open.

• Stated but unmeasured: the quality claim itself. "Smarter than most popular MoE models with a similar memory footprint" and "faster than most popular dense models" ship without a benchmark table in the announcement. No independent evaluation of Bolt 74B appears to exist yet.

• Unknown as of writing: the context window; the named licence (NVIDIA says open source without naming one in this post); the exact inference throughput at the 24GB tier; whether the release is downloadable weights only or also served behind an API; whether 2-bit is the only published precision; and how closely the architecture descends from the Nemotron 3.5 Lightning line NVIDIA has already shipped in BF16 and NVFP4 builds.

That last unknown is the one to hold onto. NVIDIA's own naming carries both numbers — "74B-A9B" and the earlier "30B-A3B" — so the two are comparable on paper: roughly 2.5× the total parameters of the Lightning model and 3× the active ones. Whether that ratio survives 2-bit compression is exactly what a first independent benchmark would settle.

Bolt E4B is the other half of the family, and it is a different product

Nemotron 3.5 Bolt E4B, shipping October 23, is not a smaller cousin of the 74B in the way the naming suggests. NVIDIA calls it "the third-generation small language model (SLM) for NVIDIA ACE," the suite it uses for in-game AI characters — a lineage the company traces through a PUBG beta for the first generation and an Aniimo preview at Gamescom for the second.

Its stated design goals are gaming-shaped rather than agent-shaped: a 4GB footprint, "fast time to first token," "a new conversational mode [that] helps keep responses concise and speech-ready," a reasoning toggle, and grounding in application state for multi-turn tool use. NVIDIA pairs the release with a test drive inside SkyrimNET, a Skyrim NPC modding platform that supports both cloud and local inference, starting the same day.

So if you arrived from the tweet expecting one model, you are getting two on different dates, in different envelopes, for different jobs: October 15 for the 24GB local agent, October 23 for the 4GB game character.

How to evaluate it without betting a roadmap on it

The awkward part of a day-one local model is that you cannot yet answer the only question that matters — does the 2-bit checkpoint hold up on your workload — and you also cannot answer the second one, whether there is a hosted route to compare against. Both gaps are temporary, and neither is a reason to plan around a benchmark table that does not exist.

The practical path is to keep the two decisions separate. Run Bolt 74B locally as an experiment in the 24GB envelope you already have, and keep the production path on whatever is callable today. That second half is where routing infrastructure earns its place: one key across 200-plus models at provider list price with 0% markup means trying a second or third candidate costs a config change rather than a second contract, and automatic failover across providers means an unproven local build never becomes the single point of failure for a live agent.

It is also the honest way to read NVIDIA's own framing. "Purpose-built for running local agents on RTX Spark" is a statement about where the model runs, not about what it scores. Treat the October 15 date as the moment a question becomes answerable — not as an answer.

What would change this article

Three things, in rough order of how much they would move the picture.

• A named licence and a public weights repo. "Fully open source including training data, weights and recipes" is the strongest claim in NVIDIA's post and the easiest to verify. Until the licence is named and the files are downloadable, that claim is a promise.

• A first independent benchmark. The interesting comparison is not Bolt 74B against a full-precision 70B-class model it was never competing with. It is Bolt 74B at 2-bit against every other build that fits in 24GB, measured on the same harness.

• A hosted endpoint. Without one there is no price to argue about, and no way to route to it. If a served version appears, the pricing comparison against the models already on your shortlist becomes concrete, and pass-through pricing becomes worth checking the same day.

Until at least one of those lands, the correct posture is interest rather than commitment. The tease was real, the dates are real, and the model is real. The evaluation is not.

Questions worth answering properly

Is Nemotron 3.5 Bolt 74B the same model as Nemotron 3.5 Lightning 30B-A3B?

No. They are different entries in the same generation with different shapes: Bolt is 74B total with 9B active, Lightning is 30B total with 3B active on NVIDIA's own naming. Bolt also ships at a 2-bit, 24GB footprint, while NVIDIA has published the Lightning line in BF16 and NVFP4 builds. Whether they share an architecture is not something the announcement says.

Can you run Bolt 74B on a 24GB desktop card, or only on an RTX Spark laptop?

NVIDIA's stated constraint is the footprint, not the machine — 24GB at 2-bit — and the named runtimes are llama.cpp, vLLM, Ollama, Unsloth and LM Studio, which are not NVIDIA-exclusive. RTX Spark is where NVIDIA is pointing the marketing, but nothing in the announcement makes it the requirement.

Does OrcaRouter host Nemotron 3.5 Bolt 74B?

Not today, and there is nothing to host yet — NVIDIA has published no served endpoint or price for it. If that changes, the model becomes callable through the same key as everything else on the catalogue, at the provider's list price with no markup added.

For a model whose failure modes are still undocumented, the useful counterpart is the record of what has already gone wrong elsewhere: the documented AI incidents and agent failures 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.