Moonshot AI Launches Kimi K2 — Beats ChatGPT in Coding & Math
Chinese artificial intelligence startup Moonshot AI has officially launched its latest open-source AI model, Kimi K2.
This new large language model designed to perform complex tasks like coding, tool use, and multi-step reasoning autonomously.
What’s new?
The new model, boasting a massive one trillion total parameters, is a Mixture-of-Experts (MoE) architecture with 32 billion active parameters per token, demonstrating significant advancements in coding capabilities, general agent tasks, and tool integration.
Kimi K2 comes in two variants: Kimi-K2-Base, a raw foundation model for researchers and builders seeking full control for fine-tuning, and Kimi-K2-Instruct, a post-trained, chat-ready “drop-in” model designed for general-purpose chat and agentic experiences.
Kimi K2: Built to Act, Not Just Think
Moonshot AI highlights Kimi K2’s strength in agentic intelligence—its ability to independently use tools, generate and run code, and carry out complex, multi-step tasks without human help.
Unlike models that only process information, Kimi K2 is built to take action. The company claims it makes advanced agent-like capabilities more open and accessible than ever before.
Benchmarks & Efficiency:
K2 achieves its high performance at different benchmarks which includes:
Agentic Coding: Achieved 65.8% accuracy on SWE-bench Verified, a challenging software engineering benchmark, matching or outperforming several proprietary models.
Live Coding: Scored 53.7% accuracy on LiveCodeBench, a realistic coding benchmark, surpassing DeepSeek-V3 and even GPT-4.1.
Mathematics & STEM: Recorded a remarkable 97.4% on MATH-500, outperforming GPT-4.1’s 92.4%.
Multilingual: Shows competitive performance on MMLU-Pro across more than 20 languages.
Availability
Kimi K2 model released under a permissive license, free for developers, with commercial pricing far below GPT‑4.1 or Claude.
News Gist
Moonshot AI has unveiled Kimi K2, a powerful and efficient open-source AI that outperforms GPT 4.1 and rivals in coding, math, and autonomous tasks.
With its innovative architecture, advanced optimizer, and low-cost access, Kimi K2 is part of China’s strategy to reclaim leadership in AI.
OpenAI Chooses Google Cloud to Power ChatGPT
OpenAI Launches ChatGPT Agent That Can Build Spreadsheets & PowerPoints
UK Switches On Isambard AI, Its Most Powerful Supercomputer Yet
Lightricks AI Now Generates Full 60‑Second Videos in Real-Time
Mistral Launches Voxtral—Open‑Source AI for Speech
