Moonshot AI Releases Kimi K3 Amid Open-Weights Discussions
Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open-weights Mixture-of-Experts (MoE) model featuring native vision capabilities and a 1-million-token context window. This release coincides with ongoing industry-wide discussions regarding the definition and implications of open-weights AI models. As the world's first open 3T-class model, Kimi K3 pushes the boundaries of what is publicly accessible, offering advanced capabilities in long-horizon coding and deep reasoning. It highlights the growing trend of releasing massive, frontier-class models with open weights, challenging the dominance of closed-source proprietary models. Kimi K3 utilizes native MXFP4 quantization-aware training from the supervised fine-tuning (SFT) stage onward, employing MXFP4 weights with MXFP8 activations for broader hardware compatibility. Developers can access the model via Moonshot AI's API, which is compatible with OpenAI and Anthropic formats.
## BACKGROUND
An "open-weights" AI model is one where the trained parameters (weights) are publicly shared, allowing developers to run and customize the model locally, unlike fully closed APIs. However, "open weights" differs from true "open source," as the training data, code, and pipeline are not always fully disclosed. Moonshot AI is a prominent AI startup known for its Kimi assistant and focus on long-context processing.