Hardware and harness news led today: NVIDIA and Microsoft pitched RTX Spark laptops as the local runtime for Windows AI agents. Mistral Large 4 chatter continued (preview pricing, October weight drops, and Ecosia walking away from Mistral toward Chinese open models). Security dominated the back half: AI-assisted intrusions against South Korean banks and a PoeLLM botnet hijacking thousands of inference servers.
Today
- NVIDIA, Microsoft kick off RTX Spark and AI agents on Windows PCs — Jensen Huang and Satya Nadella framed co-engineered PC hardware/software for on-device agents — the “local inference” counterweight to cloud-only copilots. NVIDIA Blog
- Mistral opens trillion-parameter Large 4 preview; weights due in October — Timeline and API-preview details as Europe’s flagship model moves from announcement to downloadable weights. BetaNews
- Mistral Large 4 preview: what’s confirmed and what’s hype — Sanity check on benchmark claims and preview pricing before weights land. Online Tech Tips
- Ecosia drops Mistral and eyes Chinese AI models — A European search brand switching inference suppliers is a real-world signal on cost, quality, and geopolitics of model choice. Cybernews
- Chutes and Harvard release 6.12B LLM request dataset — Large-scale production traffic dump for studying routing, latency, and real-world prompt mix (research, not a product launch). shattered.io
- Architect launches Liquid Inference — real-time auction for LLM inference — Market mechanism for spot capacity; relevant if you are optimizing token spend across providers. MarkTechPost
- PoeLLM cryptojacking hits 3,400 AI servers — Botnet hides C2 in a GitHub poem — reminder to lock down inference endpoints and GPU fleets. shattered.io
- ARTEX AI pentesting tool used in data theft attacks on South Korean financial firms — Offensive AI tooling showing up in live intrusions, not just red-team demos. The Hacker News
- CrowdStrike analyzes AI-enabled hack on South Korea’s financial sector — Vendor attribution and tradecraft notes on AI-assisted campaigns. 디지털투데이
- Stack Overflow publishes Level 3 of LLM production maturity — Confidence calibration and human handoffs as the next ops layer beyond “it answers.” news.lavx.hu
- What model weights are — and why the license decides what you can build — Primer on Apache/MIT vs research-only weights for product teams. SQ Magazine
- Open-source vs closed-source LLMs: what should you actually use? — Decision framing for teams picking between API vendors and self-hosted weights. HackerNoon
- Open-weight AI changed the race, not the rules — Macro essay on competition vs safety and compliance burdens. streamlinefeed.co.ke
From NVIDIA
- Gears of War: E-Day launches on GeForce NOW — Cloud gaming catalog update (light on core AI; included for feed completeness). NVIDIA Blog
Watch: Open-weight models
- Mistral Large 4 arrives with “Le Chonk” nickname — Culture + product story as Europe’s challenger model gets memed and measured. Inc.com
- JEV-27B-VL trending on Hugging Face — Open “decision” VL models gaining traction alongside Liquid-style edge stacks. Yahoo Finance
Watch: OpenAI · local inference narrative
- RTX Spark laptops and local 284B models (benchmark claims) — Press cycle around on-device models vs cloud GPT-class APIs — treat benchmarks as unverified until independent runs. Wccftech
Built with utilities/ai_report_build.py (24-hour lookback). No new Innermost Loop edition in this window. Edited for publication. About the AI Report.