Kimi K3 (Fully Tested): AN OPEN MODEL BEATS FABLE?!
Sponsored by Moonshot AI (Creators of Kimi K3): https://www.kimi.com/code In this video, I'll be testing the upcoming Kimi K3 model from Moonshot and showing how it performs on my KingBench benchmark, including frontend tasks, three.js tasks, SVG generation, math reasoning, long-horizon agentic work, and a hard 3D wristwatch test. -- Key Takeaways: 🚀 Kimi K3 is a massive jump over previous Kimi models and feels like one of the strongest models I’ve tested. 📊 It scores 62 out of 80 on KingBench, placing third overall behind Fable 5 and Opus 4.8. 🧠 Kimi K3 solves the hard math problem correctly with the exact answer, 20460, earning a full 10 out of 10. 🛠️ The model shines most in long-horizon agentic tasks, completing the Gemma fine-tuning and web UI task fully autonomously. 🎨 It performs very well on frontend, animation, SVG, and three.js tasks, including the elevator simulation and folding table. ⚙️ Kimi K3 uses tools extremely well, including checking browser outputs through Chrome CLI and fixing issues on its own. 💡 It does not overthink simple tasks, making it feel fast, practical, and potentially cost-efficient. 🏆 In real-world agentic work, it feels better than Opus 4.8 and GPT-5.6 Sol, though Fable 5 still feels slightly stronger overall.
