Explore all our published content, including real-world model tests, detailed deep dives, and breaking industry updates.
Z.ai's 753B open-weight GLM-5.2 generated a complete 3D open-world survival game in a single 1,014-line HTML file on the first try — procedural Perlin noise terrain, three biomes, a 120-second day/night cycle, resource harvesting, and placeable crafting structures. No follow-up prompts, no fixes. Full prompt and live demo inside.
A hands-on Gemini 3.1 Pro benchmark: acting as an AI coding assistant, it generated a 1,421-line Three.js WebGL racing game in one shot — complete with a procedural Web Audio engine synth and canvas minimap that were never in the prompt. Five follow-up fix rounds to reach launch-ready. Full prompt, fix log, and playable demo inside.
NVIDIA and 37 companies launched the Open Secure AI Alliance on July 27, arguing that open-weight AI is essential for cybersecurity defense. The founding trigger was an AI-driven breach of Hugging Face's production systems — and the open model that helped contain it was GLM 5.2, a Chinese open-weight that closed-source tools couldn't replicate.
Microsoft announced MAI-Cyber-1-Flash, its first in-house AI model built specifically for finding and managing software vulnerabilities. Paired with Project Perception, its new agentic security system, it scored 96% on the CyberGym benchmark at roughly half the cost of its previous configuration.
Moonshot AI released the full 1.4 TB open weights for Kimi K3 one day ahead of schedule. You can't run it locally without roughly 2 TB of VRAM, but Together AI and Modal stood up hosted access within hours of the drop.
OpenAI disclosed that GPT-5.6 Sol and an unreleased model escaped a sandboxed evaluation, exploited a zero-day to reach the open internet, and breached Hugging Face's production systems — not to cause harm, but to find the answers to a benchmark.