Mistral's Trillion-Parameter Drop, VS Code's Ghost Agents, and Hacked Freezers

▸ • Mistral drops Large 4, a 1T-parameter behemoth with a 1M context window, but watch out for the hosting license limits.

▸ • VS Code's new Agent Host lets your AI coding sessions keep running in the cloud even after you close your laptop.

▸ • A suspected cyber-attack is hitting supermarket freezers across Europe, proving that the IoT dystopia is officially chilling.

🌪️ Mistral Large 4 Hits the Scene

Mistral just dropped a 1T-parameter, 1M-context multimodal monster. It's downloadable, but the MIT-derivative license puts a leash on reselling it as a service, so read the fine print before you start your own API wrapper startup.

👻 VS Code's Ghost Agents

GitHub's latest VS Code update decouples AI sessions from your editor window. You can now kick off a task, close your laptop, and supervise your AI agent from your phone like a true micromanager.

🧠 Nace AI's Local Decision Maker

Nace AI open-sourced Drex Max v1, a 27B parameter decision model that runs locally on Apple Silicon. It's perfect for routing and policy checks without sending your sensitive data on a round-trip to the cloud.

📟 The $48 E-Paper Dream

A new DIY e-paper display works across two official ecosystems or entirely on your own server. It's the perfect excuse to finally build that smart home dashboard you've been procrastinating on.

🍦 Vanilla JS Strikes Back

One developer built 59 free, browser-based dev tools using strictly vanilla JS and native APIs. No React, no servers, no telemetry—just pure, unadulterated Web Crypto and Canvas APIs doing the heavy lifting locally. It's a refreshing reminder that the browser is actually a pretty capable operating system.

⚛️ October's React & AI Ecosystem Catch-up

If you blinked in October, you missed Next.js 16.4 pushing Cache Components, TypeScript 7.1 hitting beta with faster checks, and TanStack dropping five announcements. Time to update your package.json, run npm install, and pray nothing breaks.

🔄 SenseTime's Layer-Reusing AI

SenseTime open-sourced a 9B parameter model that runs a shared block of 24 decoder layers twice. It's a clever architectural trick to extend processing depth without bloating the model weights, proving that sometimes you just need to double-dip.

🐈 LongCat-2.5's Agentic Multimodality

A new 1.6T parameter model designed specifically for long-horizon agentic tasks. It uses multi-level discrete tokens to process visual and textual inputs in a unified format, making it a powerhouse for complex codebase development.

🇪🇺 Aleph Alpha's Kolibri

A 78B parameter open-weight German-English model released under Apache 2.0. With over 21% German training data, it's a strong play for European digital sovereignty and localized AI deployments.

🏗️ NGX View Builder 1.0

A visual builder for Angular that lets you drag and drop forms and dashboards, saving the result as JSON. It even includes an MCP server so your AI assistant can build screens while you sit back and watch.

🧊 The Great Freezer Hack of 2026

Supermarket freezers across Belgium and Sweden have been hit by a suspected cyber-attack. Because nothing says 'modern warfare' quite like melting everyone's frozen pizzas.

🍰 Portal's Lost Prequel Leaks

A fully playable version of the canceled Portal prequel, F-Stop, has leaked online. Time to dust off your gravity-defying puzzle skills and see what Valve left on the cutting room floor.