Kimi K2: A 2.8-Trillion-Parameter Open-Source Model with Open Weights Dropping July 27
Moonshot AI's Kimi K2 is a massive sparse mixture-of-experts model with 2.8 trillion total parameters, though only about 16 out of 896 experts are active at any time. It supports a 1 million token context window with native vision, and its architecture includes new components called Kimi Delta Attention and Attention Residuals for handling long contexts efficiently. The model is already live through the Kimi API and apps, with full open weights releasing July 27, 2026, priced at roughly $3 per million input tokens and $15 per million output tokens.
Why It Matters
Open weights on a model this large — with strong coding and agentic benchmarks — could mean you can self-host or fine-tune a frontier-class model without licensing restrictions.
🔒 Get-started steps + learning link
Subscribe to unlock the get-started steps and learning link for every item.
More from AI & Agentic Skills
- ChatGPT Can Now Read Your Medical Records — Here's How to Use It Safely
- Run Image-Generation Models Faster and Cheaper with Nunchaku 4-Bit Inference in Diffusers
- OpenAI Presence Lets Businesses Deploy Voice and Chat AI Agents Without Building from Scratch
- OpenAI's Safety Report on Long-Running AI Agents Is a Must-Read Before You Build One