Moonshot AI, the Chinese lab behind the Kimi assistant, has announced K3, a large language model with 2.8 trillion parameters and a context window that can span up to one million tokens. The company said it will fully open‑source the model weights on July 27.
Where K3 sits in the landscape
According to the Artificial Analysis intelligence index, K3 ranks third, after Fable 5 and GPT‑5.6 Sol. Observers note that Moonshot’s API pricing is roughly comparable to Anthropic’s Sonnet tier. The K3 launch follows months after DeepSeek crossed the trillion‑parameter threshold, underlining that trillion‑scale models are becoming increasingly common in China.
The real constraint: serving the model
Technically, publishing weights is straightforward for well‑resourced labs; the harder part is operating a model of this scale. Running a 2.8‑trillion‑parameter model requires vast compute capacity, specialized accelerators and substantial electricity. Moonshot and analysts warn that those runtime resources are scarce, and access to serving capacity may sell out quickly.
In practice this means that while anyone can download the K3 weights after July 27, actually deploying the model for real‑world, high‑traffic use will remain expensive and limited to organisations or states that can afford the necessary hardware and energy.
Implications
K3’s open weights reinforce a recurring pattern in the current AI race: openness of model artifacts is not by itself sufficient to democratise access. Control over compute infrastructure and power becomes a decisive factor in who can operate and monetise the largest models. Researchers will gain a valuable resource when the weights are released, but the economics and logistics of serving such a model will continue to concentrate practical capability among the best‑funded players.
