Nvidia Readies RTX Spark Software Ecosystem: 74-Billion-Parameter Model Fits in 24GB, Open-Source Release Set for October 15
SANTA CLARA, California — Nvidia on October 13 published a software ecosystem update for its RTX Spark platform, ahead of the October 16 on-sale date for RTX Spark laptops.
The centerpiece is Nemotron 3.5 Bolt 74B-A9B, a new model built for running SI agents locally and scheduled for release on October 15. Using a mixture-of-experts design with 74 billion total parameters and about 9 billion active per token, plus 2-bit quantization-aware training, the model fits into 24GB of memory. Nvidia said it will be fully open source — weights, training data and recipes — and will run at launch through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, with integrations for agent apps such as Hermes and OpenClaw. A smaller sibling, Nemotron 3.5 Bolt E4B, aimed at conversational agents and game characters, follows on October 23.
Also on October 13, Nvidia's free, open-source PAIR (Personal AI Router) moved out of beta to general availability: it discovers compatible PCs on a local network and distributes inference across them, with the release adding RTX Spark and llama.cpp support and direct model downloads from Hugging Face. Broadcast 2.3 shipped with a built-in Model Context Protocol (MCP) server, agent apps including GitHub Copilot and Perplexity gained new integrations, and a sandboxed agent execution framework built with Microsoft entered early preview.
RTX Spark is Nvidia's Windows PC platform, pairing a Grace CPU with a Blackwell GPU and up to 128GB of unified LPDDR5X memory. First laptops from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte go on sale October 16, with Microsoft's Surface Laptop Ultra starting at $2,599; compact desktop systems are slated for November.
(Source: Nvidia's October 13 announcement and specialist outlets including Tech Times and VideoCardz)
Published by GSM · USA (original Chinese edition)
Photo: 英伟达位于加州圣克拉拉的总部大楼（Endeavor），CC BY-SA 4.0（Wikimedia Commons，摄影：Coolcaesar）
