RTX Spark laptops ship Friday: a 74B agent model squeezed into 24GB runs locally

Local AI agents on the PC get a single starting point: on first boot, users see a Getting Started screen for choosing and installing an agent app. Local LLM downloads were already five times higher than all of last year, but agents still required users to assemble models, inference engines and security settings themselves. NVIDIA and Microsoft, ahead of launch, coordinated ecosystem partners so that GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy recommend an optimized local model and settings for RTX Spark and complete setup in a few clicks.

The trade-off between local agents and cloud models is handled by a router, so users can cut cloud usage while still approaching cloud-only results. Previously users either ran local models and gave up frontier intelligence or ran cloud models and paid the token cost. NVIDIA Switchyard is an open source model router library that lets agents route queries between local and cloud LLMs; Hermes will add Switchyard later in October, and Perplexity Computer and GitHub Copilot have integrated their own routing solutions.

A 74B-parameter model built for local agents is compressed into 24GB of memory while retaining quality. Models with a similar memory footprint were previously either dense and slower or mixture-of-experts and less intelligent. Nemotron 3.5 Bolt 74B-A9B 24GB is a 74B-parameter, 9B-active-parameter mixture-of-experts model compressed with 2-bit quantization and quantization-aware training, releasing Oct. 15, runnable through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, with training data, weights and recipes fully open source.

The report covers software updates NVIDIA, Microsoft and ecosystem partners announced ahead of the RTX Spark launch, for developers and creators running local agents on a PC. The stated next steps are the Oct. 16 laptop launch and the Oct. 15 release of the Nemotron 3.5 Bolt 74B model.

NVIDIA says Nemotron 3.5 Bolt 74B is faster than most popular dense models and smarter than most popular mixture-of-experts models with a similar memory footprint, but gives no independent test conditions. Adobe's up to 2x performance gains, the 2x faster DLSS 5 model and up to 1.4x frame-rate boost are vendor comparisons whose actual effect depends on the application and hardware.