A 2.8-trillion-parameter model released Friday by Beijing-based Moonshot AI matches or beats top-tier proprietary systems from OpenAI and Anthropic on key benchmarks, and it arrives as open weights that any developer can download and modify. VentureBeat rewrites a long-held assumption, that open-source AI trails closed-source frontier models by at least six months. Kimi K3 closes that gap almost entirely, landing just ahead of the 2026 World Artificial Intelligence Conference in Shanghai where Chinese President Xi Jinping called AI development "a symphony of global cooperation."
What K3 actually does
Kimi K3 is 75% larger than DeepSeek's V4 Pro, with a 1-million-token context window, native visual understanding, and an always-on reasoning mode Moonshot calls "thinking mode." Reuters Its architecture uses two internally developed innovations -- Kimi Delta Attention and Attention Residuals -- both previously published as open research on GitHub. The benchmark results are the headline. K3 scored 1,687 on GDPval-AA v2 (real-world tasks across 44 occupations), placing third behind only Claude Fable 5 Max and GPT-5.6 Sol Max.
On BrowseComp, a benchmark for long-horizon information seeking, it hit a state-of-the-art 91.2 out of 100. In four of eight real-world task automation benchmarks -- including Automation Bench, SpreadsheetBench 2, and BrowseComp -- K3 ranked first.
Forbes "The model alone is no longer the product," Perplexity CEO Aravind Srinivas told CNBC last week. "It is the harness, the orchestration system that puts the model inside a very capable harness and pairs the model with a lot of tools." CNBC
Market impact and pricing
The Moonshot news hit competitors hard. Shares of Zhipu and MiniMax plunged 27.7% and 16.5% respectively in Hong Kong trading.
Z.ai, which released a new model in June, dropped 28.4%.
K3's API pricing sits at $3 per million input tokens and $15 per million output tokens, with cached input tokens at $0.30 per million. That's roughly half the cost of OpenAI's GPT-5.6 Sol, according to a Bank of America research report Friday. A promotional rebate through August 12 offers up to 30% back in vouchers for API credits of $1,000 or more.
Moonshot itself is on a growth trajectory that predates K3. The company reported annual recurring revenue above $200 million in April, driven by subscriptions for Kimi and its other AI services.
It raised roughly $2 billion in its latest round, valuing it at more than $20 billion.
The politics of open weights
Full model weights are scheduled for release on July 27, making K3 available for anyone to self-host or fine-tune. That open-weight strategy carries geopolitical weight.
Chinese firms including DeepSeek, Alibaba, Tencent, and Baidu have all released open-source models, but none at this parameter count. Bloomberg U.S. lawmakers are debating whether American companies should use Chinese open-source models, a conversation that Patrick Moorhead, CEO of Moor Insights and Strategy, called "an over-reaction shockingly similar to the DeepSeek panic." Moorhead told CNBC the technology will "accelerate and grow the inference market faster than without."
Moonshot hasn't disclosed what hardware it used to train K3, but the startup is a partner with Huawei. The company faces U.S. export controls limiting access to advanced Nvidia chips, a constraint that hasn't stopped it from producing what Arena CEO Anastasios Angelopoulos called "the single biggest release of the year."
The comeback story
Founded in 2023 by Tsinghua graduate Yang Zhilin (Ph.D. Carnegie Mellon, 2019), Moonshot was one of China's most prominent AI startups before DeepSeek's R1 model disrupted the market in early 2025.
Kimi had ranked third in monthly active users in China but slid to seventh. K3 is the culmination of a pivot to open-source models that began with Kimi K2 in July 2025.
Yang's former adviser Russ Salakhutdinov, a former director of AI research at Apple, wrote: "It feels like just yesterday Zhilin was graduating from my lab at CMU."
