Moonshot AI releases Kimi-K3 model weights (2.8T params) and open-sources infrastructure; performance near-parity with OpenAI/Anthropic frontier models.

Moonshot AI has released the weights for Kimi K3, enabling anyone with contemporary AI GPU infrastructure to run and monetize the model with minimal restrictions. This represents a significant competitive challenge to leading closed-source providers, particularly Anthropic and OpenAI.

In benchmarks published by Moonshot, Kimi K3 outperforms earlier generations of Claude and GPT while closely matching the latest frontier models—Claude Fable and GPT-5.6 Sol—while appearing to be 2-3x cheaper to operate. With cached queries, the cost advantage extends to 10x. Moonshot's technical documentation substantiates these claims, disclosing the specific testing methodologies employed.

Input token pricing demonstrates the efficiency gains. Moonshot charges $3 per million tokens for Kimi K3, compared to $10/1M for Fable and $5/1M for Sol at standard, non-cached rates. However, Kimi K3's caching mechanism reportedly achieves a 90% hit ratio for coding tasks, reducing effective input costs to $0.30/1M. Output token economics follow a similar pattern.

Several architectural factors drive this efficiency. Kimi K3 uses MXFP4 precision for weights and MXFP8 for input activation—low-precision data types requiring substantially less VRAM. Of its 2.8 trillion parameters, only 104.2 billion activate at any given time. The model employs a fixed-size state mechanism called Kimi Delta Attention instead of conventional key-value caching, reducing memory and computational overhead. Its sparse mixture-of-experts layer activates only 16 experts of 896 total, further minimizing inference costs.

Notably, Moonshot's testing employed Nvidia's H20 GPUs, a relatively modest chip by current standards lacking native MX floating-point support. This optimization strategy could render Kimi K3 particularly amenable to lower-end hardware, though running it on specialized silicon like Nvidia Blackwell might substantially improve performance beyond current benchmarks.

The open-weight release carries strategic implications: it enables direct competition from entities with adequate GPU infrastructure while reinforcing the perception that high-performance models can be freely available and dramatically cheaper than proprietary alternatives. However, "open-weight" denotes only unrestricted model weights—Moonshot retains exclusive knowledge of training processes and datasets.