NVIDIA posted the software stack for RTX Spark on October 13, two days before the laptops: Bolt, a 74 billion parameter open mixture of experts due Thursday, PAIR out of beta with llama.cpp, and Copilot plus other agent apps with Spark-specific setup. Google DeepMind and Singapore's GovTech described an on-device Gemma that scores scam conversations without sending the chat to the cloud. AWS put Grok 4.7 on Amazon Bedrock with US and global inference profiles. NVIDIA committed $250 million over five years to UK agentic safety research and joined the Earthshot Prize as a Global Alliance Founding Partner.
In the news
RTX Spark software
NVIDIA's October 13 post is the software side of RTX Spark, two days before laptops become available on Friday, October 16. The headline model is Nemotron 3.5 Bolt 74B-A9B, a 74 billion parameter mixture of experts with 9 billion active, packed into 24 GB with 2-bit quantization-aware training. NVIDIA says the open weights, data, and recipes land Thursday, October 15, and that the model will run through llama.cpp, vLLM, Ollama, Unsloth, LM Studio, Hermes, and OpenClaw. Bolt is not on Hugging Face yet.
NVIDIA PAIR, the personal AI router that discovers machines on a local network, leaves beta in the same post. The general-availability build adds llama.cpp, Hugging Face model downloads, and RTX Spark discovery on Windows, Linux, and macOS. GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi, and WorkBuddy add Spark-tuned setup plus OpenShell and Microsoft Execution Containers for sandboxing. NVIDIA quotes up to twice the performance for Adobe Photoshop and Premiere GPU work, still rolling out, and a fivefold jump in local LLM downloads versus last year. Those figures are the company's.1
On-device Gemma
Google DeepMind and Singapore's Government Technology Agency posted on October 14 about a privacy-preserving scam detector built on Gemma. A lightweight classifier screens inbound messages, and a fine-tuned Gemma then scores the multi-turn thread. The model runs entirely on-device with user consent, and DeepMind says neither Google nor the Singapore government can read conversations through the deployment.
Training used anonymised scam-report corpora inside a GovTech-managed GCP sovereign sandbox, not citizen PII. DeepMind, citing The Straits Times, says scams are more than half of Singapore's reported crime, more than S$4 billion in losses since 2019, and that social-engineering now accounts for about 80 percent of those losses. The work was presented at Singapore International Cyber Week 2026. App rollout is still phased.2
Grok 4.7 on Bedrock
AWS's October 12 machine learning post puts Grok 4.7 on Amazon Bedrock. Requests go through cross-Region inference profiles us.xai.grok-4.7 for US residency and global.xai.grok-4.7 for commercial Regions, on the Converse API and on OpenAI-compatible Chat Completions and Responses. SpaceXAI lists a 500,000-token context window, text and image in, text out, tool calling, structured output, and a reasoning-effort knob at low, medium, high, and xhigh.
AWS built a tool-calling agent and ran 72 evaluations across three search-not-lookup tasks, four effort levels, and six repeats. It reports 67 of 72 correct, 18 of 18 at xhigh, and 15 of 18 at low using about an eighth of the output tokens. That study is AWS's own synthetic eval. This is cloud availability of a model that already exists, not a new Grok drop. Grok 4.8 has not shipped.3
NVIDIA in the UK
NVIDIA's October 14 UK ecosystem post commits $250 million over five years to UK agentic safety and security research. The press release frames that as a compute donation, $50 million a year, delivered through UK-based NVIDIA Cloud Partners, with a portion supporting BOLD, SOFAIR, and Imperial College London. NVIDIA is also joining the Earthshot Prize as a Global Alliance Founding Partner, offering compute and technical expertise to climate and sustainability finalists.
The same brief names Fuse Energy on B300 Blackwell Ultra, Locai Labs' open Nemotron-Personas-UK dataset of one million synthetic personas, OneAdvanced NHS primary-care triage, and Danaher's claim of up to eight times faster affinity-reagent discovery. Those partner numbers are vendor-reported. Jensen Huang's GTC Berlin keynote is Wednesday, October 21, at 11:00 CEST.4
