Kimi to Release Open Model 'K3'

Chinese AI startup Kimi has announced the release of a new open-weight model called 'K3.' An open-weight model is a format in which the model's weights—numerical data that underpins AI thinking—are publicly disclosed, allowing researchers and developers to freely use and modify them. The complete weight data is scheduled to be released by July 27, 2025.

K3 is a multimodal model capable of processing not only text but also images and other content. It has 2.8 trillion parameters and a context length (the maximum amount of text that can be processed at once) of up to 1 million tokens. A context length of 1 million tokens is comparable to reading several full-length novels at once, making it suitable for long-document analysis and complex task applications.

According to Kimi's own benchmark evaluation, K3 demonstrates performance comparable to Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol. Meanwhile, against Opus 4.8 and GLM 5.2, K3 showed results that sometimes surpass them by significant margins, and the company highlights its competitiveness approaching top-tier models. However, it should be noted that this benchmark is conducted by Kimi itself, and independent third-party verification has not yet been performed.

Notable changes are also evident in pricing. K3 has seen a significant price increase compared to the previous generation, signaling a departure from the 'ultra-low-cost' approach that has become synonymous with Chinese AI models. Over the past few years, Chinese AI companies have leveraged competitive pricing as a strategic weapon. The price increase for K3 can be understood as a shift in focus from price competition to differentiation based on performance and quality.

In context, the open-weight model market has seen large-scale models like Meta's Llama series widely adopted in the developer community, with increasing enterprise adoption as well. Kimi's introduction of K3 into this landscape demonstrates renewed competitive strength in the open-model domain from Chinese developers.

Meanwhile, Claude Fable 5 and GPT 5.6 Sol, which the company cited as benchmarks, are proprietary (closed-source) models currently regarded as top-tier. If K3 achieves performance comparable to these as an open-weight model, it could expand options for enterprises and research institutions to perform sophisticated processing in-house without relying on expensive APIs.

Following the July 27 weight release, comprehensive benchmark validation by independent researchers and developers is expected to accelerate. How well the company's self-reported performance claims are reproduced will be a critical inflection point for K3's market adoption. Combined with the shift in pricing strategy, this could also influence how the broader open-model market views costs.