@adamstewartmarketing: Moonshot just dropped an open-weight AI model that’s beating Claude Opus at coding. It’s Kimi K3.1 Pro from Moonshot AI, and the full weights are free to download from November 4. The numbers: • SWE Marathon: Kimi 51.7, Claude Opus 5.5 46.3, GPT-6 Astra 42.5 (Moonshot’s own benchmark) • Artificial Analysis independently ranks it 4th out of 215 models • One team gave Kimi and Opus the exact same coding task. Both passed all 96 hidden tests. Kimi was slower, but cost about half as much (at low reasoning effort) The part almost nobody is talking about: before releasing the weights, Moonshot stripped hacking data out of its training. So it’s worse at writing exploits than Opus, on purpose. Once a model’s weights are public, there’s no taking them back. The catch: at 2.8 trillion parameters, you won’t be running this on your laptop. You’d need some serious GPUs, or about four maxed-out Mac Studios. But you can plug it into Claude Code today and pay about half of what Opus charges per token.

Adam Stewart | AI & Marketing

Open In TikTok:

Region: AU

Sunday 25 October 2026 12:09:49 GMT