Moonshot AI has introduced Kimi K3, a new frontier model positioned around long-context reasoning, coding, vision input and planned open-weight availability. Official materials describe a 2.8-trillion-parameter architecture, while independent tracking from Artificial Analysis and reporting from the Associated Press...
Moonshot AI has introduced Kimi K3, a new AI model that the company describes as a 2.8-trillion-parameter system built for long-context reasoning, coding and multimodal use.
In its official Kimi K3 announcement, Moonshot AI says the model uses a 2.8T-parameter architecture and introduces technical components including Kimi Delta Attention and Attention Residuals. The company also says Kimi K3 supports native vision input and a 1 million-token context window.
Moonshot’s positioning for Kimi K3 emphasizes long-horizon coding and complex task execution rather than only short prompt-and-response interactions. The company says full weights are planned for release by July 27, 2026, which would make the model more accessible for researchers and developers who want to inspect, deploy or adapt it outside a closed hosted environment.
The official Kimi API Platform documentation also lists Kimi K3 as available for developer access through Moonshot’s API, giving developers a route to test the model without waiting for the planned full-weight release.
Artificial Analysis, an independent AI model benchmarking and pricing service, lists Kimi K3 as released in July 2026. Its model page ranks Kimi K3 at number four out of 187 models on the Artificial Analysis Intelligence Index.
Artificial Analysis also lists Kimi K3 as supporting both text and image input, with a 1 million-token context window. The service reports pricing of $3 per million input tokens and $15 per million output tokens.
Those figures matter because long-context models can become costly when used for software repositories, document review or multi-file debugging. A 1 million-token context window gives developers and enterprises room to process large bodies of material, but practical adoption will still depend on latency, reliability, accuracy and cost under real workloads.
The Associated Press reports that Moonshot’s newest Kimi K3 model appears to be catching up to leading Claude and ChatGPT models. AP also reports that Kimi K3 topped Arena’s front-end coding ranking, a result likely to attract attention from developers evaluating AI systems for web application work.
Front-end coding benchmarks are only one slice of model performance. They do not prove that a model is best across security, architecture, backend engineering or production maintenance. But they are relevant because modern AI coding assistants are increasingly judged not just on whether they can write snippets, but on whether they can build, revise and debug interactive software.
Moonshot’s own Kimi K3 materials frame the model around long-horizon coding. Combined with AP’s reporting on Arena’s front-end coding ranking, that suggests Moonshot is targeting one of the most commercially important uses for advanced language models: software development.
Moonshot’s plan to release full weights is one of the most significant parts of the Kimi K3 announcement. Open or inspectable model weights can allow outside researchers to evaluate model behavior more directly, and they can give companies more flexibility over deployment, fine-tuning and data controls.
However, the value of that openness will depend on the final license terms, the completeness of the released assets and the practical requirements for running the model. A 2.8T-parameter model is not a lightweight system, and even open weights may require substantial infrastructure or specialized deployment strategies.
AP frames Kimi K3 as part of a broader wave of Chinese AI models challenging leading U.S. systems. Moonshot AI’s announcement arrives in a market where model providers are competing on reasoning ability, coding, context length, multimodal input, price and deployment flexibility.
For developers, the near-term question is whether Kimi K3’s benchmark results and long-context design translate into dependable day-to-day performance. For researchers and enterprises, the planned full-weight release could make Kimi K3 a model to watch closely, particularly if Moonshot follows through with terms that support meaningful external evaluation and deployment.
The available sources support a cautious conclusion: Kimi K3 is a notable new entry in the frontier-model race, with official claims around scale, long context, vision and coding, independent benchmark visibility from Artificial Analysis, and early mainstream attention from the Associated Press. Its long-term significance will depend on real-world performance, openness details and developer adoption.
Moonshot AI has introduced Kimi K3, a new AI model that the company describes as a 2.8 trillion parameter system built for long context reasoning, coding and multimodal use.
What Moonshot AI announced In its official Kimi K3 announcement, Moonshot AI says the model uses a 2.8T parameter architecture and introduces technical components including Kimi Delta Attention and Attention Residuals.
The company also says Kimi K3 supports native vision input and a 1 million token context window.
Continue reading
