Moonshot AI just dropped what might be the most consequential open-source AI model of the year. Kimi K3, a 2.8-trillion-parameter multimodal beast, is performing on par with the best publicly available models in agentic coding tasks, according to Dean Ball, OpenAI’s head of strategic futures.
What K3 actually does
Kimi K3 uses a mixture-of-experts architecture, activating 16 out of 896 total experts for any given task. Instead of firing up the entire 2.8-trillion-parameter engine for every query, it selectively engages only the specialists it needs.
K3 comes with a 1-million-token context window, which means it can process roughly the equivalent of several full-length novels in a single pass. It also ships with native vision capabilities, so it can interpret images and visual data without bolting on a separate model.
On benchmarks, K3 scored 67.5 on DeepSWE, a benchmark that tests sustained software engineering ability. It posted a raw pass rate of 77.8 on ProgramBench. Ball specifically highlighted K3’s strength in agent programming scenarios — the multi-step, tool-using workflows where an AI model needs to plan, execute, and adapt over extended coding sessions.
