Beijing-based Moonshot AI released Kimi K3 on 16 July 2026, a 2.8-trillion-parameter open Mixture-of-Experts foundation model built on the company's Kimi Delta Attention and Attention Residuals architecture that activates 16 of 896 experts per token. It is the largest open-weight release from a Chinese lab to date and marks Moonshot's third major foundation-model release inside 12 months.
Delta Attention And Extreme Sparsity
Kimi K3 pushes Moonshot's sparse-expert design further than the trillion-parameter Kimi K2 released last summer. Activating just 16 of 896 experts per token cuts inference cost sharply relative to comparable dense frontier models, while Delta Attention and Attention Residuals aim to improve long-context reasoning and agentic tool use. The release continues a wave of Chinese open releases challenging U.S. proprietary systems on efficiency and cost.
An Aggressive Release Cadence
Moonshot released Kimi K2 in July 2025, K2.5 with the MoonViT vision encoder in January 2026 and Kimi K2 Thinking in November 2025. Kimi K3 lands two weeks after Anthropic's Claude Sonnet 5 and one week after OpenAI's GPT-5.6 Luna, Sol and Terra, all within a 10-day window that saw multiple frontier models ship.
Company Context
Founded in March 2023 by Tsinghua University alumni Yang Zhilin, Zhou Xinyu and Wu Yuxin, Moonshot is one of China's so-called "AI Tigers." Alibaba took a 36% stake and led a US$1 billion round at a US$2.5 billion valuation in 2024, Tencent joined a US$300 million round later that year, and the company is reportedly weighing a Hong Kong IPO in 2026. Moonshot's Kimi assistant has tens of millions of monthly professional users and its Mooncake serving platform handles roughly 100 billion tokens a day. The company is also building on its long-context research strengths.
What K3 Signals
Moonshot's approach - trillion-scale open weights, extreme MoE sparsity and a public release schedule closer to weeks than quarters - keeps pressure on U.S. frontier labs. It also raises the bar for the open-source ecosystem: at 2.8 trillion parameters, K3 is far larger than most Llama or Qwen releases, but its sparsity keeps single-node inference feasible for well-funded developers.
Reporting based on coverage from LLM Stats, CNBC, Nature, South China Morning Post and Moonshot AI.