China-based Moonshot AI shared the loads of its highly anticipated new model, Kimi K3, free of charge. This breakthrough allows anyone with current artificial intelligence GPU hardware to run the model on their own systems.

This decision of the company is considered as an important step against leading players in the field such as OpenAI and Anthropic. According to the test results shared by Moonshot AI, the K3 outperforms the previous generation models and has a performance that can compete with its current rivals.

High efficiency and low cost operation

Some K3 offers significant advantages over its competitors, especially in terms of operational costs. While the entry cost of the model is determined as 3 dollars per million tokens, this number can decrease to 0.30 dollars in reasonable usage scenarios, thanks to the caching structure.

The basis for the efficient operation of the model lies in the use of low-precision MXFP4 and MXFP8 data types. In the model, which has a total of 2.8 trillion parameters, only 104.2 billion parameters are activated during each process.

Moonshot AI states that it used lower-level hardware compared to today’s standards, such as Nvidia H20, in its testing processes. This indicates that the model can work efficiently on a wider range of hardware.

Instead of traditional key-value storage, a fixed-size state manager called Some Delta Attention is preferred. This architecture both reduces VRAM usage and contributes to overall efficiency by shortening the process time.

Competition stability in the sector is changing

Presenting the K3 as an open-loaded model paves the way for developers and companies to become direct competitors to big players by establishing their own infrastructure. This development creates a valuable alternative, especially for cost-oriented projects.

The introduction of the open-loaded model does not change the fact that the training process and data set are still the exclusive property of Moonshot AI. Still, this step proves that free software can operate at much lower costs while performing comparable to its commercial rivals.

The model is expected to show even higher performance when running on modern hardware with MXFP support, such as Blackwell. Official tests to be carried out in the future will reveal the real impact of these optimizations on the hardware side.

How do you think this breakthrough by Moonshot AI will affect the market share of giants such as OpenAI and Anthropic?