Kimi K3 | China's Free AI Models Are Getting INSANELY Good
Kimi K3 just topped the frontend code leaderboard for free, while Chinese firms still smuggle banned Nvidia chips at huge risk Kimi K3, Moonshot AI's new flagship, just took the number one spot on Arena's Frontend Code leaderboard, beating paid frontier models from Anthropic and OpenAI, while a Chinese firm's own tax filings confessed to $92 million in hardware built for banned GPUs. Both things are true at once, and the reason exposes why export controls stop chips at the border but do nothing once a model's weights hit Hugging Face. This video breaks down the actual numbers behind Kimi K3: a 2.8-trillion-parameter Mixture-of-Experts model that only activates 16 of 896 experts per token, reads a million-token context window, scores 91.2 on BrowseComp and 42.0 on SWE Marathon, and prices input tokens at 30 cents to $3 per million. It covers why Moonshot built it this way, native MXFP4 quantization designed to run on standard enterprise hardware instead of a hoard of restricted GPUs, and what that means once the full weights go public on Hugging Face. It also looks at the hardware side: Huawei's Ascend 950PR chip claims, the Council on Foreign Relations' read that the compute gap with Nvidia is widening, the Super Micro smuggling indictment, and the panic-buying of two million chips the moment H200 export licenses loosened. Then it connects this to the bigger shift MIT Technology Review documented, Alibaba's Qwen family overtaking Meta's Llama in downloads, eight of the top ten open-source models now Chinese, and the single afternoon DeepSeek's R1 wiped 17% off Nvidia's stock. For builders and the AI-curious trying to understand what China's open-weight strategy actually means for the AI agents and models you'll be building with next. Chapters: 0:00 Number one and still smuggling chips 0:34 How good is insanely good really 1:43 Why risk prison for chips anyway 3:02 Frontier power without frontier chips 4:16 Files don't stop at customs 5:19 The takeover already happened 6:24 One file that crashed Nvidia 7:19 The war nobody declared Tools & resources mentioned: - Kimi K3 (Moonshot AI): https://huggingface.co/moonshotai - Hugging Face: https://huggingface.co - Qwen (Alibaba): https://huggingface.co/Qwen - DeepSeek R1: https://huggingface.co/deepseek-ai - Huawei Ascend 950PR - Nvidia H200 About The Stack The Stack helps you build with AI. Each video takes one tool, model, or workflow and shows how it works in a few focused minutes, with the real benchmarks and real costs. We go deep on Claude Code and Cursor for AI coding, AI agents and MCP servers, the open-source AI tools and GitHub repos most people miss, RAG and vector search, fine-tuning, and running local LLMs on your own machine with Ollama and LM Studio. We compare models like ChatGPT and Claude, test AI automation with Zapier, Make, and n8n, and flag the tools that actually ship. Subscribe for new breakdowns: https://www.youtube.com/@the-stack-ai?sub_confirmation=1 #kimik3 #aiagents #opensourceai #chinaai #machinelearning
