Summary
Fireship’s Code Report digs into Kimi K3.1, the new three-trillion-parameter open-weight model from Moonshot AI. The model first appeared anonymously on Chatbot Arena under the name Kibitz before Moonshot confirmed it, with full weights promised for download on November 4th.
The video examines why Kimi’s performance is surprising given Anthropic’s claim that it now blocks 90 to 99 percent of distillation attacks from Chinese labs, and OpenAI’s similar efforts. Kimi K3.1 lands within a point of Opus 5.5 on Terminal Bench 4.0, and Artificial Analysis independently measured 67 percent, the highest score it has recorded. The host also notes that the test harness can swing results significantly, and covers Kimi’s strong showing on a long-horizon software benchmark.
Other topics include Moonshot’s decision to strip cybersecurity data from training, which cut its internal exploit-test success from 90.5 to 81 percent, and the context of a recent US ban on DeepSeek for federal civilian agencies. The video closes with a tongue-in-cheek 3D horse test in Three.js, comparing Kimi’s visually impressive but poorly animated result against Claude, GPT-6 Astra, and Google’s output.
šŗ Source: Fireship Ā· Published October 26, 2026
š·ļø Format: News Analysis
