Moonshot unveils Kimi K2 Thinking, a GPT-5 rival for agents

Moonshot AI has announced Kimi K2 Thinking, an open-weight language model designed to solve long-horizon problems, research the web and chain together actions with tools. The company says the model outperforms closed systems such as GPT-5 and Claude Sonnet 4.5 on tests including Humanity’s Last Exam and BrowseComp.

The news matters for one specific reason: open models are no longer competing only on conversation, coding or cost. Kimi K2 Thinking is aiming at one of the areas where closed AI providers have built much of their advantage: agents that can reason through many steps and act on external tools.

One trillion parameters, but not all work at once

Kimi K2 Thinking has one trillion total parameters, though it activates 32 billion for each operation. This architecture is known as a mixture of experts: the model is made up of many specialized blocks and selects only some of them to answer each request.

The distinction matters. The total number of parameters gives an indication of the system’s accumulated capabilities, but active parameters account for much of the cost and speed of use. Moonshot is therefore seeking to combine a massive knowledge base with more efficient inference than a comparably sized dense model can provide.

Its weights are publicly available on Hugging Face. That allows researchers, developers and companies to download, study and run the model on their own infrastructure, provided they have the necessary hardware. It does not mean using it is easy or cheap: a model of this scale still requires servers with considerable memory capacity.

The bet is on agents that do not stop after two steps

Kimi K2 Thinking’s main stated capability is its ability to make between 200 and 300 sequential tool calls. A call could involve a web search, a database query, code execution or reading a document. The model observes the result, decides what to do next and continues until it completes the task.

That behavior is at the core of so-called AI agents. A conventional chatbot answers a question; an agent, by contrast, can break an assignment into subtasks, search for information, cross-check it, use a calculator or modify code files. The challenge is not just making one tool call, but maintaining the goal and staying on track after dozens of decisions.

Moonshot presents the model as built for this kind of reasoning and extended tool use. Put simply, the goal is for the system to complete multistep tasks instead of limiting itself to producing a single response.

Strong results, with one important caveat

Moonshot says Kimi K2 Thinking outperforms GPT-5 and Claude Sonnet 4.5 on Humanity’s Last Exam and BrowseComp, in the configurations it evaluated. Humanity’s Last Exam consists of difficult questions from academic disciplines; BrowseComp measures the ability to find information online that requires multiple searches and intermediate deductions.

That is a meaningful comparison, but it does not by itself settle which model is best in every situation. Results depend on the tools allowed, the step budget, the data available and how responses are evaluated. Moreover, strong performance on an exam does not guarantee that an agent will be reliable when managing real-world business processes.

If confirmed by reproducible, independent evaluations, the signal would be significant for the open ecosystem. Kimi K2 Thinking aims to narrow the gap with proprietary systems in a category that is particularly valuable for automating research, document analysis, technical support and software development.

The challenge shifts from accessing the model to operating it well

For companies, open weights offer three practical advantages: greater control over data, the ability to adapt the model to internal workflows and less dependence on a single commercial API. In return, they must take on the infrastructure, the security of connected tools and error evaluation before delegating important tasks.

The next test for Kimi K2 Thinking will not be a benchmark table. It will be whether the model can sustain that long-horizon reasoning outside prepared environments, with incomplete documents, changing websites and ambiguous objectives. If it can, Moonshot will have given the open community something more valuable than a large model: a real alternative for building advanced agents without handing the entire intelligence layer to a closed provider.