What happened

• Moonshot AI released Kimi K3.1, described as a three-trillion-parameter open-weight model, with a downloadable release announced for November 4.
• The video connects its launch to Kibitz, an anonymous model that appeared on Chatbot Arena and performed at frontier level. The narrator jokingly suggests its name hinted at Kimi’s involvement.

Benchmark claims and caveats

• Kimi K3.1 reportedly scored within one point of Opus 5.5 on Terminal-Bench 4.0. Artificial Analysis independently measured it at 67%, its highest result among tested models.
• It also reportedly led on SWE-MARATHON, a benchmark for long, complex software tasks.
• The narrator cautions that benchmark results can vary substantially depending on the evaluation harness.

Distillation and security

• The video discusses Anthropic’s accusations that Moonshot and other Chinese AI companies used large-scale distillation of Claude outputs. It presents the broader debate over whether distillation is improper copying or a way to learn from AI systems trained on human-created material.
• The narrator questions how Kimi’s performance fits with claims that distillation attacks are being blocked, while noting that benchmark scores alone do not establish how a model was trained.
• Moonshot reportedly removed some cybersecurity data from Kimi’s training set. An internal, unfiltered version allegedly solved 90.5% of Moonshot’s exploit tests, compared with 81% for the released model.
• Despite those limitations, Kimi reportedly ranked highly on Artificial Analysis’s cyber index; the video notes that some competing models declined the test prompts.

The “Thoroughbench” 3D-horse test

The narrator’s informal test asked models to build a realistic, interactive 3D horse using Three.js.

| Model | Score | |---|---:| | Google Argon | 0.69 | | GPT-6 Astra | 5.21 | | Opus | 6.39 | | Kimi K3.1 | 4.20 |

• Kimi produced the most visually appealing horse in the sample, with realistic coloring.
• Its animation was weak: it appeared to warp the mesh rather than use a skeletal rig.
• Claude’s less attractive horse used a mesh with about 95,000 vertices and 35 independently animated bones, producing more convincing movement.
• Verdict: Kimi’s result was impressive visually, but it scored below the other frontier models overall.

Sponsor mention

• The video promotes Mobbin, a library of 600,000 UI screens with an MCP server. Coding agents can use it to study real app flows, gather design references, and adapt those patterns to a project. The sponsor offers a free trial and a 20% discount through the video’s link.