Beijing-based Moonshot AI unveiled Kimi K3 on 17 July, a 2.8 trillion-parameter open-weight model the company says matches or beats several leading American systems, at a fraction of the price.
The release landed on the eve of Shanghai's World Artificial Intelligence Conference.
Moonshot, backed by Alibaba Group Holding and Tencent Holdings, built K3 as the flagship of a three-tier lineup that also includes K2.7 Code, a specialised coding model, and K2.6, a general-purpose option.
What K3 actually does
K3 is the largest open-weight model released by anyone to date at 2.8 trillion parameters, more than double the size of its predecessor, Kimi K2, which shipped at 1 trillion parameters in July 2025.
Context windows across the three-model lineup run to 256,000 tokens or above, and K3 offers the full 1-million-token version, enough to process an entire codebase or a lengthy document in a single pass.
Moonshot claims K3 performs competitively against Anthropic's Fable 5 and beats OpenAI's GPT-5.6 on several published benchmarks, though those figures come from the company itself.
An independent ranking on Arena.AI reportedly put K3 at the top of current model performance.
The model shows particular strength in front-end coding, a category Moonshot is also targeting with its open-source Kimi Code tool.
Cheaper pricing and strategy
K3 costs US$3 per million input tokens and $15 per output, against something closer to $50 for comparable American frontier models.
It's expensive by Chinese standards - z.ai's GLM-5.2 charges US$4.40 per million output tokens and DeepSeek's V4 charges just US$0.87 - but still dramatically cheaper than its U.S. rivals.
Context caching is automatic, with no cache ID, TTL, or extra parameter required.
Full model weights are scheduled for public release on 27 July, meaning developers will be able to download, study and self-host K3 without paying API fees to anyone.
The company's pivot to open-source began with Kimi K2 in July 2025 and accelerated with K2.5 in January 2026.
Training a model of this scale requires enormous computational resources and months of preparation.
The company's own chart positions K3 against other open-source frontier models, showing it well ahead of DeepSeek's 1.6 trillion parameters, Xiaomi's 1.02 trillion and Alibaba's 397 billion.
About the company
Yang Zhilin, Zhou Xinyu and Wu Yuxin founded Moonshot in March 2023; Yang previously worked as a professor at Tsinghua University.
Moonshot raised $2 billion in May at a valuation exceeding $20 billion, with backers including Alibaba, Tencent, Meituan and HSG.
The company's revenue exceeded $200 million, according to a statement from its financial advisor, though as a private company Moonshot does not publish audited financials.
