Last Week in AI #334 - Kimi K2.5 & Code, Genie 3, OpenClaw & Moltbook
Last Week in AI Last Week in AI
China's Moonshot just open-sourced Kimi K2.5, a free AI that reads text, images and video. It's built so swarms of AI agents can team up and actually finish tasks.
Based on reporting by Last Week in AI, Last Week in AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Moonshot AI, the Beijing-based lab that's been quietly climbing the open-weights leaderboard, just dropped Kimi K2.5. It's a natively multimodal model, meaning it wasn't bolted together from separate vision and language systems after the fact. Moonshot trained it from scratch on 15 trillion tokens of mixed text, images, and video, which is a staggering pile of data even by 2024-25 standards.
What's more interesting than the raw scale is the framing. Moonshot isn't just selling K2.5 as a smarter chatbot. They're pitching it as agent infrastructure, leaning hard into what they call "agent swarm" orchestration, where multiple instances of the model coordinate on a task instead of one model grinding through everything solo. That's a meaningfully different pitch than the usual benchmark-chasing releases we've seen out of Chinese labs this year.
And the fact that it's open source matters here. While OpenAI, Anthropic, and Google keep their flagship weights locked behind APIs, Moonshot is handing this out for anyone to run, fine-tune, or build a coding agent on top of, which is exactly what accompanied this release: a dedicated coding agent built on K2.5. That's the pattern with Chinese AI labs lately, DeepSeek included, they release fewer press statements and more actual weights.
This was one thread in a busier-than-usual week for AI news, alongside Google's Genie 3 world model, the OpenClaw project, and something called Moltbook that also made the rounds. But K2.5 stands out because it's not a research demo, it's a model people can download today and start plugging into real agent pipelines.
My take — AI-written commentary, not fact-checked reporting
I'll say the obvious thing nobody wants to hear: the most consequential open models right now aren't coming out of San Francisco, they're coming out of Beijing, and that gap is going to keep widening as long as Western labs treat openness as a marketing afterthought. Betting the house on API-gated frontier models while Moonshot hands out 15-trillion-token multimodal weights for free is a strategic error dressed up as a safety argument.
Read more about this at: Last Week in AI