Kimi released K3, a multimodal open-weight model with 2.8 trillion parameters and one million token context window. The model performs close to Claude Fable 5 and GPT 5.6 Sol in internal benchmarks, surpassing Opus 4.8 and GLM 5.2 in several test categories. The company plans to release full weights by July 27.
K3 represents a notable shift in Kimi's pricing strategy. The model costs significantly more than its predecessor, signaling an industry-wide move away from the ultra-low-cost AI models that defined the Chinese market. This pricing increase reflects the computational demands of training larger, more capable models and suggests consolidation around quality-focused competitors rather than race-to-bottom pricing.
The model's scale and performance metrics place it in the upper tier of open-weight options available today. With one million token context, K3 handles extended documents and complex multi-turn conversations without degradation. The multimodal architecture processes both text and images, expanding use cases beyond pure language tasks.
Kimi's release comes amid broader industry trends. As foundation models become more capable, the economics of production shift. Training larger models demands substantial compute resources, making sustainable pricing necessary. Chinese AI developers who built market share through aggressive pricing now face pressure to demonstrate quality parity with Western models like Claude and GPT variants.
The open-weight release strategy differs from closed API approaches. By distributing full model weights, Kimi enables researchers and developers to fine-tune K3 for specific tasks, potentially offsetting the higher base cost through customization value. This approach also builds developer loyalty in a market increasingly fragmented between open and proprietary systems.
K3's positioning suggests the era of rock-bottom AI pricing has ended. Companies must now compete on capability, reliability, and ecosystem support rather than