Kimmy K3.1 Review: Benchmarks, Coding Tests, and a 3D Horse Showdown

Moonshot AI's new open-weight model Kimmy K3.1 has arrived with benchmark numbers that rival frontier closed models, raising questions about how it was trained and what it can actually do. This episode examines the model's claimed performance, the controversy around distillation and safety controls, and puts it through a practical 3D coding test using Three.js.

■ Kimmy K3.1 Release and Distillation Controversy
- Anthropic's allegations that Chinese AI companies harvested Claude outputs for training
- Whether blocked distillation attacks conflict with Kimmy's strong benchmark results
- Terminal-bench and long-horizon software task scores, plus independent verification
■ Safety and Cyber Security Trade-offs
- How Moonshot removed cyber security data from its training set
- Internal versus released model results on exploit tests, and the US government ban on DeepSeek
- Rankings on Artificial Analysis' new cyber index
■ 3D Horse Benchmark Test with Three.js
- A head-to-head comparison of Kimmy K3.1, Claude, GPT-6 Astra, and Gemini Argon
- Mesh quality, bone-based skeletal animation, and final scores
■ Sponsored Segment: Mobin
- MCP server integration and its library of real UI screens for coding agents

This is aimed at developers and AI practitioners who want a grounded read on a new open-weight model rather than headline benchmark numbers. Viewers will come away understanding the context behind Kimmy K3.1's release, the limits of its reported scores, and how it performs on a concrete, reproducible coding task.