Kimi K3 Explained: Agent Swarm Collaboration, Deep Thinking & 3D Game Generation 75

Kimi K3 is the new flagship model released by Moonshot AI on July 16, 2026, and is currently the open-source foundation model with the largest parameter scale in the world. Kimi K3 adopts a MoE architecture, supports a 1 million token context window, and is natively multimodal, with outstanding capabilities in coding and reasoning. The model is open-weight, allowing developers to freely download, deploy, or call the API.

Core Parameters

Total parameters of 2.8 trillion, adopting a MoE architecture, with a context window of 1 million tokens, supporting text and images. The architecture uses the proprietary Kimi Delta Attention (KDA) hybrid linear attention + Attention Residuals (AttnRes). The complete weights are expected to be open-sourced on July 27 under the Modified MIT license.

Key Features

The long context has been extended from K2's 256K to 1 million tokens, capable of processing an entire novel or large codebase in one go. The KDA mechanism improves long-text decoding speed by up to 6.3 times.

In terms of coding, it focuses on Agentic Coding, capable of understanding large codebases, calling tools, executing tests, and continuously iterating. It supports Agent Swarm, which can coordinate up to 300 sub-agents in parallel to handle complex tasks.

Being natively multimodal, it can make comprehensive judgments by combining images, code, runtime logs, and test results, making it particularly practical for scenarios such as game development, front-end engineering, and CAD design.

The Attention Residuals technique adds less than 2% extra computational cost while improving training efficiency by approximately 25%.

Performance

According to official claims, its overall intelligence level is close to that of the world's leading closed-source models. It ranks among the top open-source models in benchmarks such as SWE-bench Verified, LiveCodeBench, Tau2, and aiME, with some items approaching top-tier international closed-source models. Overall, it is slightly weaker than Claude Fable 5 and GPT-5.6 Sol, but can align with Claude Opus 4.8 and GPT-5.5.

Pricing

It has been launched on kimi.com, Kimi Work, and Kimi Code. The API is compatible with the OpenAI SDK, priced at $3 per million tokens for input, $15 per million tokens for output, and $0.30 per million tokens for cached input. The open-source weights will be available for download, deployment, secondary development, and fine-tuning on July 27.

Kimi K3 marks Moonshot AI, following DeepSeek, as another leading Chinese AI company to open up its most advanced model to the world in the form of open weights, further narrowing the performance gap between China and the United States in cutting-edge AI models.