Kimi K3 Raises Pressure on OpenAI and Anthropic
By Suad Seferi ·
Chinese artificial intelligence company Moonshot AI has released Kimi K3, a 2.8-trillion-parameter model designed for coding, research and complex tasks that can run for extended periods with limited human supervision. The model is available through Kimi’s website, desktop applications, coding tools and API. Moonshot says it will release the full model weights on July 27, when developers should be able to examine and deploy K3 more independently. Kimi K3 arrives at a moment when Chinese AI companies are becoming increasingly difficult for their American competitors to dismiss. OpenAI and Anthropic still lead many of the most closely watched evaluations, but models from Moonshot, DeepSeek, Z.ai and MiniMax are narrowing the distance while competing aggressively on access and price. The market noticed. The Wall Street Journal reported that Moonshot’s announcement intensified Friday’s selloff in chip stocks, as investors weighed what cheaper and increasingly capable Chinese models could mean for companies benefiting from the enormous spending behind the AI boom. K3’s headline number is its 2.8 trillion parameters, but the model does not use all of them for every request. It relies on a mixture-of-experts architecture that activates 16 out of 896 specialised components at a time. The approach allows Moonshot to build a much larger model without using the full computing capacity for every task. It also comes with a one-million-token context window and native support for visual information. Moonshot is positioning the model for software engineering, document analysis, scientific research and other work that requires an AI system to retain information and operate across many steps. In demonstrations published by the company, Kimi K3 worked on GPU optimisation, built a small compiler, contributed to chip design and completed multi-stage research tasks. Those examples are ambitious, although they were selected and presented by Moonshot. Independent testing will be needed to determine how reliably the model performs outside controlled conditions. Moonshot has been more restrained than some AI companies in describing where K3 stands against its main competitors. It acknowledges that the model’s overall performance still trails OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5. The company nevertheless claims that K3 competes closely in several coding and knowledge-work evaluations, while outperforming other tested models in selected tasks. Benchmark comparisons remain difficult because different models are often tested with different tools, settings and levels of reasoning effort. Moonshot also disclosed limitations that could affect practical use. K3 may become unstable if an application fails to preserve its previous reasoning history. It can also act too proactively when instructions are unclear, making decisions that the user did not explicitly request. That behaviour is not a minor concern for a model designed to operate for long periods. The more access an AI system has to code, files and external tools, the more costly an incorrect assumption can become. Pricing will be another part of Moonshot’s challenge to established providers. API access costs $0.30 per million cached input tokens, $3 per million uncached input tokens and $15 per million output tokens. Developers can use the model now, although the promised release of its weights has not yet happened. That distinction matters. Kimi K3 is being presented as an open model, but developers will have to wait until July 27 to see what Moonshot releases, under which licence and with what technical documentation. The company says a fuller report covering the model’s architecture, training and evaluations will follow. K3’s arrival coincides with the World Artificial Intelligence Conference in Shanghai, where China is presenting new models, computing systems and proposals for international AI governance. The timing places Moonshot’s release within a wider effort to show that China can compete across the entire AI stack despite US restrictions on access to advanced chips. The release does not mean that Kimi K3 has overtaken the leading American models. Moonshot itself is not making that claim. What it does show is that the gap is no longer wide enough to make Chinese models an afterthought. The more important judgment will come after the weights and technical report are released, when researchers and developers can test whether Kimi K3’s performance holds up beyond Moonshot’s own demonstrations.