NVIDIA Nemotron 3.5 Bolt Puts 74B-Parameter Agents on Laptops

NVIDIA has released Nemotron 3.5 Bolt, a 74-billion-parameter open model purpose-built for agentic AI, compressed into a 24GB footprint so it can run entirely on a single high-end PC or a new RTX Spark laptop. The release, announced on October 15, 2026, arrives alongside Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model aimed at specialized always-on agents. Together, the pair marks NVIDIA's most direct push yet to move AI agents off cloud APIs and onto local hardware — the same day RTX Spark laptops go on sale.

According to the company's announcement, both models are fully open under the OpenMDW-1.1 license, which permits commercial use, and the weights were posted to Hugging Face on the day of release. As reported by Blockchain.News, NVIDIA framed the three-model lineup as covering the full range of agent deployments, from enterprise servers to desktops and edge devices, without forcing developers to trade capability for deployability.

A 74-billion-parameter model squeezed into 24 gigabytes

Nemotron 3.5 Bolt 74B-A9B uses a mixture-of-experts design: 74 billion parameters in total, with only about 9 billion active per token. Per NVIDIA's account, native 2-bit quantization — applied with quantization-aware training so quality is preserved through compression — packs Nemotron 3.5 Bolt into 24GB of memory, small enough to run alongside other applications on a single machine. The company's test hardware includes a PC with a GeForce RTX 5090 and a DGX Spark, and the model is also targeted at RTX Spark laptops.

Nemotron 3.5 Bolt supports up to 256,000 tokens of context, and the Hugging Face model card lists text, image, and video inputs for the 24GB checkpoint. According to Tech Times, the answer to the "frontier-class on consumer hardware" question is precisely this decoupling of total knowledge from per-token compute: the sparse mixture-of-experts design keeps inference cost near that of a 9-billion-parameter model while the full 74 billion parameters remain available. NVIDIA says the model is intended for everyday local agents that search the web, use tools, and write or debug code.

At launch, Nemotron 3.5 Bolt runs through llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, and per MACCUS, the weights, training data, and recipes are all public — developers can fine-tune the model and customize their own agents on top of it. NVIDIA has also wired the release directly into agent frameworks: Hermes and OpenClaw are among the first to support it. A still smaller model, Nemotron 3.5 Bolt E4B 4GB, is scheduled to arrive on October 23 for edge and embedded deployments.

Super VL: the 120-billion-parameter agent specialist

Nemotron 3.5 Super VL is the larger sibling, and the one NVIDIA positions closest to the "always-on agent" ideal. It combines a hybrid Mamba-2/Transformer mixture-of-experts backbone with a C-RADIO vision encoder, and according to The Neuron, it handles text alongside images, document pages, charts, GUI screenshots, video frames, and PDFs — the visual surface an agent needs to operate a computer the way a human does. Supported deployments offer up to one million tokens of context.

NVIDIA described Super VL as its fastest open agentic model in its class, citing Artificial Analysis Intelligence Index leaderboard results showing up to five times faster throughput than other models in its class. The company also claims the model matches the accuracy of its largest reasoning model at roughly one-fifth the parameter count, and targets instruction following, tool use, multi-step business workflows, and visual understanding. Independent benchmarks of Nemotron 3.5 Bolt will tell the full story — as reviewers have noted, the quantization, speed, and intelligence claims are NVIDIA's own, and hands-on testing against the shipped weights is still arriving.

Super VL was built in collaboration with the Nemotron Coalition, and a third model remains on the way: Bolt E4B 4GB, designed for physical and conversational agents on GeForce RTX, Jetson, and Spark hardware, arrives October 23. The Nemotron 3.5 Bolt 74B checkpoint itself was compressed from Super VL, per the technical account.

RTX Spark laptops go on sale — with an agent software stack

The hardware side of the story launched the same week. According to Review71, RTX Spark laptops built on a Grace CPU plus Blackwell GPU, with up to 128GB of unified memory, go on sale October 16 — one day after the release of Nemotron 3.5 Bolt, the 74B model NVIDIA purpose-built for agentic AI on the new hardware. NVIDIA's first-boot setup will let users choose an agent app, and tools including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi, and WorkBuddy are being prepared to recommend an optimized local model and configuration, reducing the command-line work that has kept local models harder to try than a cloud chatbot.

NVIDIA also moved PAIR — its free, open-source router for distributing AI work across compatible computers on a local network — out of beta. The full release adds llama.cpp support, direct model downloads from Hugging Face, and RTX Spark support. Broadcast 2.3 shipped with a built-in Model Context Protocol server, and a sandboxed agent execution framework built with Microsoft became available in early preview.

Security is part of the stack, too. NVIDIA's OpenShell and Microsoft's Execution Containers add system-level containment, credential protection, and monitoring for agents that need access to files and apps. According to Review71, OpenClaw and GitHub Copilot already support the container primitives, with Hermes integration due later in October. The protection these controls offer depends on configured policies and how each agent uses them — a point that echoes across this week's agent-governance news, from open-source frameworks like Builder.io's Agent-Native release to credential systems like ServiceNow's AI Passport with NVIDIA as its design partner.

Why on-device agents matter for the agent economy

The significance of Nemotron 3.5 Bolt is not another leaderboard score; it is where the model runs. With Nemotron 3.5 Bolt, an agent model that lives on a laptop means always-on personal agents without routing every step through a cloud API bill, and documents never have to leave the machine to be processed. For professional firms handling client materials, that keeps sensitive documents local — a governance property no cloud contract can replicate perfectly. Per Derrington Collaborative AI, NVIDIA says local LLM downloads are already five times higher than all of last year, and Bolt 74B is purpose-built for that shift.

The open license matters as much as the quantization. With weights, training data, and recipes public, smaller shops can build and fine-tune custom agents instead of renting them from a model provider — the same democratization logic that open agent frameworks have been pushing all year. Whether Nemotron 3.5 Bolt earns its place in production stacks will depend on independent benchmarks against the shipped weights, and on how the sandboxed execution layer holds up in real workflows. But the direction is unambiguous: the agent economy is coming home, one local model at a time.