Moonshot’s Kimi K3 challenges Anthropic’s frontier AI lead
Kimi K3, the world’s first open 3T-class model, outperforms Claude Opus 4.8 in controlled tests, coinciding with a high-level US-China summit and eased chip export restrictions.
Chinese artificial intelligence start-up Moonshot is preparing to launch Kimi K3, a large language model that financial markets and technology analysts view as a significant challenge to American dominance in frontier AI. The release of the model, which is anticipated to exceed the performance of Anthropic’s Claude Opus 4.8, underscores a rapidly narrowing capability gap between the United States and China in advanced artificial intelligence.
Kimi K3 is distinguished as the world’s first open 3T-class model, featuring 2.8 trillion parameters. The architecture introduces native multimodal vision capabilities and a substantial 1-million-token context window. To support these features, the model utilises specific architectural innovations, including Kimi Delta Attention and Attention Residuals, designed to enhance processing efficiency and reasoning capabilities.
In rigorous performance evaluations conducted within identical sandbox environments, Kimi K3 demonstrated superior results in GPU kernel optimisation tests. The model substantially outperformed Anthropic’s Claude Opus 4.8, as well as OpenAI’s GPT 5.6 Sol and GPT 5.5. While it performed competitively with Claude Fable 5, that model was evaluated by a third party and may incorporate fallback behaviours, making the direct comparison with Opus 4.8 a key metric for current market sentiment.
The launch occurs against a backdrop of shifting geopolitical and economic dynamics. It follows a recent two-day summit in Beijing between US President Donald Trump and Chinese President Xi Jinping, which addressed trade, artificial intelligence, and security issues. During the summit, US stock markets rallied, with the Dow Jones Industrial Average gaining 0.8% and Nvidia shares surging more than 2% after the US approved the sale of H200 chips to Chinese firms.
To manage the computational demands of expert-parallel training and inference, deployment recommendations suggest using supernode configurations with 64 or more accelerators. Open Frontier Intelligence has contributed a vLLM implementation to address challenges posed by Kimi Delta Attention regarding conventional prefix caching. The model is currently accessible via Kimi.com, Kimi Work, Kimi Code, and the Kimi API, marking a pivotal moment in the global AI race.
