Kimi K3 (low)

Kimi K3 (low) is a specific configuration of the Kimi K3 flagship model developed by Moonshot AI. Operating at a lower reasoning effort level, this version is designed to provide the core intelligence of the K3 architecture while prioritizing speed and cost-efficiency for tasks that require less intensive chain-of-thought reasoning. It remains part of the 2.8-trillion-parameter class, utilizing a sparse Mixture-of-Experts (MoE) architecture that balances massive knowledge capacity with efficient inference.

Kimi K3 is built on the Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architectures. KDA is a hybrid linear attention mechanism that allows the model to maintain high expressiveness over its 1-million-token context window without the quadratic computational growth typical of standard transformers. Attention Residuals further improve deep network training by allowing layers to selectively retrieve representations from earlier stages of the model, which is particularly effective in MoE setups with high expert counts.

Technical Configuration

The "low" designation refers to the model's reasoning effort setting. Unlike standard language models that produce immediate responses, Kimi K3 is a reasoning-centric system with "always-on" internal thinking. In its low-effort mode, the model limits the depth of its internal chain-of-thought, making it suitable for faster conversational interactions, basic coding tasks, and document retrieval where maximum planning is not required. The model still leverages its native vision capabilities, allowing it to process and reason across both text and image inputs within the same context.

Capabilities and Performance

With a total of 2.8 trillion parameters and approximately 104 billion active parameters per token, Kimi K3 (low) excels in long-horizon tasks such as navigating large code repositories and complex knowledge work. Its MoE structure features 896 total experts, with 16 experts activated per token via a Stable LatentMoE framework. This high degree of specialization allows the model to maintain state-of-the-art performance in specialized domains like programming and mathematical reasoning while operating within a massive context window of over one million tokens.