China’s Moonshot AI Releases Kimi K3, a 2.8 Trillion Parameter Open Model

11:15 · July 29, 2026

Kimi K3 combines 2.8 trillion parameters with native vision understanding, a 1 million token context window, and an always-on ‘thinking mode’ that keeps the model reasoning through a problem before producing an answer. Under the hood, Moonshot built the model around two architectural innovations it calls Kimi Delta Attention, a hybrid linear-attention mechanism designed to keep long-context inference efficient, and a technique the company refers to as Attention Residuals. Rather than keeping the model closed or licensing it selectively, Moonshot made the full weights available for unrestricted download alongside the infrastructure code used to train and run it, including its attention kernels and agent environment tooling. Founder Yang Zhilin has said the company sees openness as a way to grow its user base rather than a threat to its business, a bet that stands in contrast to how most Western frontier labs have handled their most capable models. Independent benchmarks cited by outside evaluators show Kimi K3 performing roughly on par with top proprietary systems from Anthropic and OpenAI on many tasks, though it still trails the very best closed models at the frontier. The release lands as Moonshot separately closes a large funding round, underscoring how central Kimi K3 has become to the company’s story.