LinkedIn·Thursday, 16 July 2026·16 Jul 2026

Here we go!! 🤯 18 months ago I called DeepSeek-R1 a Sputnik moment. Today the sequel dropped. Kimi K3 (🇨🇳) launched, and it's closer to…

Wouter Janssen

CEO & Co-Founder Tekst | The hardest working AI in business.

Here we go!! 🤯 18 months ago I called DeepSeek-R1 a Sputnik moment. Today the sequel dropped. Kimi K3 (🇨🇳) launched, and it's closer to the absolute frontier than any open model has ever been: - 2.8 trillion parameters → the world's first OPEN 3T-class model - 1 million token context window (that's ~10 books in one prompt) - Weights fully released by July 27. Open. For everyone. 🚨 Frontier AI was supposed to be a closed club. OpenAI, Anthropic, Google: billion-dollar training runs, and you only get access through their API. The weights? Locked away. Moonshot AI (the lab behind Kimi) released K3 and just said "here, take the whole thing." And it's not just benchmarks. In one test, K3 did in 2 hours what normally takes an experienced researcher 1 to 2 weeks: reading 20+ scientific papers, writing 3,000+ lines of code, and delivering the analysis. Is it the best model in the world? No. Moonshot admits it still trails Claude Fable 5 and GPT 5.6. But that gap is just a matter of time. What does this mean? - Open vs closed is now measured in months, not years (my bet: on par by end of Q4) - "Only US labs can play at the frontier" is officially dead - The value moves up the stack: whoever turns raw intelligence into working enterprise processes wins TLDR: You don't rent the frontier anymore. You download it. My take: Intelligence should be open. Chapeau, Moonshot. 👏 👏 👏 Full tech blog: https://lnkd.in/egMMrU4r