The Chinese AI company Moonshot AI has announced Kimi K3, a large language model with 2.8 trillion parameters that will be released as an open model later this month. With this move, the company is taking a new step in the competition between open and closed AI models. According to Moonshot, K3 is the first open model to reach the threshold of nearly three trillion parameters.
Kimi K3 was developed for complex tasks such as software development, research, and agent-based AI workflows. The model features a context window of one million tokens, enabling it to process large amounts of code or documents in a single session. Upon its launch, the model will be available through Moonshot AI’s own services, while the model weights will be released to developers and researchers by July 27 at the latest.
With K3, Moonshot AI is explicitly targeting the high end of the AI market, where OpenAI and Anthropic currently set the standard with their closed models. The company acknowledges that Kimi K3 still lags behind models such as GPT-5.6 Sol and Claude Fable 5 in overall performance, but states that the model now achieves comparable results on various benchmarks and outperforms other open models.
The announcement underscores a broader trend in the AI market. Whereas the most powerful models were long available exclusively as closed commercial services, open models that fall into the same performance category are appearing with increasing frequency. This enables companies and research institutions to run and customize advanced AI models themselves.
Designed for complex AI tasks
Moonshot positions Kimi K3 primarily as a model for long-term, complex tasks. According to the company, it can independently analyze large software projects, write programming code, control terminal tools, and carry out extensive research tasks. In addition, the model supports native image processing, allowing text and images to be combined within a single workflow.
Moonshot cites examples such as optimizing GPU kernels, developing a GPU compiler, building interactive software, and automating scientific analyses. These demonstrations are primarily intended to show that K3 is suitable for agent-based applications in which an AI works independently on a task over an extended period of time.
New architecture
Under the hood, Moonshot is introducing several architectural changes. These include Kimi Delta Attention and an updated Mixture-of-Experts setup. According to the company, these changes deliver significantly higher computational efficiency than the previous generation of Kimi models. This allows for greater performance to be extracted from the same amount of training and inference capacity.