Moonshot AI, an Alibaba-backed artificial-intelligence startup, has open-sourced its new large language model, Kimi K2. The company and early users say the system approaches GPT-4-level instruction following while remaining free to download and deploy. Kimi K2 employs a mixture-of-experts architecture with roughly one trillion parameters and can handle up to 128,000 tokens of context, allowing it to process long documents and code bases. The Unsloth developer group has already released a 1.8-bit version that shrinks the original 1.1-terabyte weights to 245 gigabytes—an 80 % reduction—so the model can run on high-end local hardware. Cloud providers and tool makers moved quickly to integrate the model. Groq reports inference speeds of about 185 tokens per second, while Anycoder, NetMind and OpenRouter have added Kimi K2 endpoints for rapid application development and agent frameworks. The rapid uptake underscores the demand for high-performance, open-source alternatives to proprietary AI systems.