Moonshot Kimi K3: 2.8T open-weight model, 1M-token context

Moonshot unveiled Kimi K3 in an announcement this week. The model contains 2.8 trillion parameters and is offered as open-weight, making it the first open-source system to approach the three-trillion-parameter class. Kimi K3 supports a one-million-token context window and includes native visual understanding. Moonshot presented the model for long-horizon coding, advanced reasoning and complex knowledge work.

Kimi K3 is built on a hybrid linear attention mechanism called Kimi Delta Attention (KDA) and incorporates a technique the company calls Attention Residuals to improve computational efficiency. Moonshot highlighted two architectural upgrades in K3 that are intended to lower compute cost per token and enable longer, stateful coding tasks with reduced human oversight.

Moonshot published comparative performance notes showing Kimi K3 performing competitively with some leading closed-source systems. In one set of benchmarks the model ranked behind a top closed model, and in tests focused on GPU kernel optimization it outperformed several others. Independent evaluations cited by Moonshot placed Kimi K3 near the top of recent model rankings and reported comparable results to high-end closed models on complex, multi-step tasks.

As an open-weight model, Kimi K3’s parameters are available for download. That allows developers and businesses to run, adapt and customize the model on their own infrastructure rather than relying solely on cloud-hosted closed-source services. Moonshot reported that Kimi-family models have led open-source scale for much of the past year.

The launch comes as Chinese AI companies have accelerated model releases and narrowed performance gaps with some US developers. Several domestic firms have introduced larger models in recent months. Moonshot has backing from Alibaba and Tencent and has been seeking fresh funding ahead of a potential Hong Kong listing.

Market reaction in Hong Kong trading was mixed. Shares of two rival AI firms fell about 27.7% and 16.5% on the day of the announcement.

Moonshot framed Kimi K3 as part of an effort to expand the scale and accessibility of open-source AI, noting scale, extended context handling and lower per-token compute costs as its technical priorities.