Moonshot AI Unveils Kimi K3, the World's Largest Open-Weight AI Model

Moonshot AI, a Chinese artificial intelligence laboratory, has released Kimi K3, an open-weight model with 2.8 trillion parameters, which the company claims is the largest of its kind publicly available.

The model, part of the Kimi family, supports a context window of one million tokens and natively processes both images and text. It employs a mixture-of-experts (MoE) architecture with 896 expert networks, of which only 16 are activated per query. This design routes tasks to specialised sub-networks, improving computational efficiency.

Moonshot AI also introduced two architectural innovations: Kimi delta attention and attention residuals. These techniques are designed to help information persist across the model's layers, leading to a reported 2.5 times improvement in scaling efficiency compared to its predecessor, Kimi K2. This means the model can more effectively convert additional compute and training data into enhanced capabilities.

In terms of performance, Moonshot AI's benchmarks indicate that Kimi K3 is competitive with leading closed-source AI systems. On coding benchmarks, it ranked first or second in most categories, including top placements on ProgramBench and SWE-Marathon. It also performed strongly on agentic benchmarks, which assess a model's ability to autonomously complete multi-step tasks, trailing only other top-tier models.

Notably, in front-end coding evaluations, Kimi K3 rose from 18th place with its predecessor to first overall, excelling in areas such as branding, data analysis, and simulations. In head-to-head comparisons, human evaluators preferred its outputs approximately 76 percent of the time, compared to around 58 percent for other leading models.

Beyond coding, Kimi K3 demonstrated improvements in general text and writing tasks, climbing from 38th to 9th on a combined leaderboard. It reached the top ten in creative writing, coding, and instruction-following, and ranked first in three professional domains: physical and social science, legal and government work, and medicine and healthcare. On an internal editorial-writing benchmark, it reportedly became the first open-weight model to top the leaderboard.

On indices meant to reflect real-world economically useful tasks, Kimi K3 placed just below the top closed-source models and ahead of several other prominent systems. On a composite index of nine challenging evaluations, including Terminal-Bench and Humanity's Last Exam, it came third, closely behind the leaders.

To demonstrate practical capability, Moonshot AI shared examples where Kimi K3 autonomously completed complex tasks: building playable browser-based games (including 3D exploration games and a Game Boy Advance emulator), generating animated motion graphics, editing videos from dozens of source clips, and even designing a functional computer chip using open-source chip design tools over a multi-hour autonomous run.

The release of Kimi K3 as an open-weight model means that researchers and developers can access and build upon the underlying parameters, potentially accelerating AI development and increasing transparency. However, it is important to note that open-weight models differ from fully open-source models, as licences and usage terms may apply.

Moonshot AI's announcement underscores the rapid evolution of large-scale AI models and the growing capability of open-weight releases to challenge the performance of proprietary systems. It should be noted that these benchmark results are based on the company's own evaluations and have not been independently verified.