NVIDIA just fired the loudest shot yet in the race to put serious AI on your desk. The company’s Nemotron 3.5 Bolt 74B, a 74-billion-parameter model purpose-built for AI agents, lands this Thursday, October 15 — squeezed into just 24GB of memory and released fully open source, including training data, weights, and recipes. It headlines a launch-week software blitz ahead of RTX Spark laptops going on sale Friday, October 16, and it changes the economics of running AI agents: the model lives on your machine, so there is no per-token cloud bill.
The announcement came in NVIDIA’s official RTX Spark launch-week blog post, published October 13. NVIDIA’s free personal AI router PAIR also moves out of beta today, and agent apps — GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi, and WorkBuddy — ship updates that make local agents easier to set up and secure.
The model, in numbers
Nemotron 3.5 Bolt 74B-A9B 24GB is a mixture-of-experts design: 74 billion parameters in total, with only 9 billion active per query. NVIDIA packed it into 24GB using 2-bit quantization with advanced quantization-aware training to preserve quality. The company’s stated target is the optimal balance between speed and intelligence — NVIDIA says it aims to be faster than most popular dense models while smarter than most popular MoE models with a similar memory footprint.
Those are NVIDIA’s own claims, and independent benchmarks will tell the full story once the weights land. But two facts are already verifiable and matter more than any leaderboard: the model is fully open source — weights, training data, and recipes — and at launch it runs through llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, with direct access inside agent frameworks like Hermes and OpenClaw.
- Release date: Thursday, October 15, 2026
- Architecture: 74B total / 9B active parameters, mixture-of-experts, 2-bit quantization
- Footprint: 24GB — laptop territory
- License: fully open source (weights, training data, recipes)
- Runs on: llama.cpp, vLLM, Ollama, Unsloth, LM Studio; Hermes and OpenClaw agents
PAIR graduates: your whole network becomes one AI computer
The sleeper story is NVIDIA PAIR, the free, open-source personal AI router moving out of beta today. PAIR auto-discovers compatible PCs on your local network, installs models and inference engines, and spreads AI workloads across every available system. The full release adds llama.cpp support, direct Hugging Face downloads, customizable inference settings, and RTX Spark support on Windows, Linux, and macOS — and the community already contributed 10+ features in the first month. For client work, that means turning a dusty office full of existing PCs into a shared AI compute pool with zero new hardware spend.
The RTX Spark launch-week stack
The model and PAIR land inside a bigger push. RTX Spark laptops go on sale Friday, October 16 — the joint NVIDIA and Microsoft announcement from October 7 set the date — with up to 128GB of unified memory and 1 petaflop of AI compute. Also shipping: agent-app security updates (NVIDIA OpenShell and Microsoft Execution Containers), NVIDIA Switchyard — an open-source router that sends queries to local or cloud models to minimize cloud spend — the 4GB Nemotron 3.5 Bolt E4B for game characters (October 23), and NVIDIA AI playbooks on build.nvidia.com (October 16).
The demand signal is striking: NVIDIA says local LLM downloads are already 5x higher than all of last year. The shift from cloud APIs to local models is not a forecast anymore.
Watch
Watch NVIDIA Developer’s official walkthrough of PAIR in action — a Hermes agent running five subagents, routed across local machines:
The money angle: three ways to cash in on the local-AI wave
Every computing shift minted a service business. Local agents are no different — except the pitch writes itself: your data never leaves your building, and you stop paying per token. Three plays: sell “local AI agent setups” to small businesses (setup fee + monthly retainer, the managed-IT playbook applied to agents); cut your own cloud bill by routing routine work to a local 74B model and reserving frontier APIs for hard queries; and fine-tune the open weights for a vertical like legal or medical intake and sell the specialized agent as a product.
What to watch next
Three dates: October 15 (Bolt 74B weights drop — watch independent benchmarks), October 16 (RTX Spark laptops on sale, plus NVIDIA’s AI playbooks), and October 23 (Bolt E4B). The honest caveat: quantization and speed claims are NVIDIA’s own until third-party testing lands. Test before you bet a client project on it.
FAQ
When does NVIDIA Nemotron 3.5 Bolt 74B release?
Thursday, October 15, 2026. NVIDIA announced the date in its October 13 RTX Spark launch-week blog post.
Is Nemotron 3.5 Bolt 74B really open source?
Yes. Like other Nemotron models, NVIDIA says Bolt 74B ships fully open source — including training data, weights, and recipes — so developers can fine-tune and customize their own agents.
What hardware do you need to run it?
The model is compressed to 24GB using 2-bit quantization, so it fits on high-memory consumer hardware and RTX Spark laptops (up to 128GB unified memory). It runs through llama.cpp, vLLM, Ollama, Unsloth, and LM Studio.
What is NVIDIA PAIR?
A free, open-source personal AI router that discovers compatible PCs on your local network and distributes AI workloads across them. It moved out of beta on October 13, 2026, adding llama.cpp support, direct Hugging Face downloads, and RTX Spark support. It works on Windows, Linux, and macOS.
When do RTX Spark laptops go on sale?
Friday, October 16, 2026, per the joint NVIDIA and Microsoft announcement from October 7.
Sources
- All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark — official NVIDIA blog, October 13, 2026 (model specs, PAIR release, launch-week software)
- NVIDIA PAIR: Hermes 5-Subagent Demo — NVIDIA Developer official YouTube channel