Loading...
Loading...
Moonshot open sourced Kimi K3's 2.8 trillion parameters along with their attention kernels, Nvidia and Microsoft formed a security alliance, and Microsoft quietly built a cyber agent that scores 96 percent.
Ten days after I wrote about Kimi K3's API, the weights landed. Modified MIT license, full technical report, and something unusual: they open sourced the attention implementation too.
Flash DA, their high-efficiency attention kernel, open sourcedWithin a day it spread across OpenRouter, Nebius and other US inference platforms. The local LLM crowd is quick to note that a 2.8T system needs serious hardware to run fast, so your gaming PC won't be hosting it anytime soon. But inference providers absolutely will, and competition does what competition does to prices.
Every open flagship release quietly reprices the entire market downward.
Same week, Nvidia and Microsoft launched the Open Secure AI Alliance: Hugging Face, IBM, Cloudflare, Dell, Palantir, Cognition and others building inspectable, open tooling for AI cyber defense. The timing wasn't subtle. This landed days after OpenAI disclosed that its own evaluation agents chained a zero-day and stolen credentials to touch Hugging Face's production servers.
Microsoft also introduced Project Perception, an agentic security system coordinating red, blue and green teams of specialized agents. Their new MAI-Cyber-1-Flash model scored 96% on CyberGym while nearly halving costs, and reportedly handles most vulnerability tasks inside their MDASH multi-agent system.
They're the same story wearing different clothes. Open weights mean offensive capability spreads to everyone, which means defensive tooling has to spread just as fast. The alliance's argument is exactly that: attackers already have frontier access, so defenders need tools they can inspect and improve, not black boxes.
Reading about an agent escaping its sandbox on Monday and tightening my own permissions by Friday probably says everything about where my head is at.
Cheers, Yassen