Model Releases
The Dawn of the 2.8T Era: Moonshot AI Unveils Kimi K3
Moonshot AI's Kimi K3 arrives as a massive 2.8T parameter MoE powerhouse, redefining long-context multimodal intelligence and agentic coding.
July 16, 2026
Model ReleaseKimi K3
Introduction: A New Milestone in Open Intelligence
On July 16, 2026, the landscape of artificial intelligence shifted once again. Moonshot AI has officially released Kimi K3, a monumental multimodal model that marks a historic turning point for the open-weights community. While the industry has long been dominated by closed-source giants, K3 represents a direct challenge to the status quo, offering frontier-level reasoning and multimodal capabilities in a package designed for massive scale.
Kimi K3 is not just another incremental update; it is a paradigm shift. By combining a massive parameter count with groundbreaking architectural optimizations, Moonshot AI has created a model that bridges the gap between accessible open-source development and the ultra-high-performance proprietary models like Claude Fable 5 and GPT 5.6 Sol. For developers, this means the ability to build sophisticated, long-horizon agentic workflows without being locked into a single provider's ecosystem.
- Release Date: July 16, 2026
- Status: Milestone Open-Weights Model
- Primary Focus: Multimodal reasoning and long-context agentic workflows
Architecture: 2.8T Parameters and Kimi Delta Attention
At the heart of Kimi K3 lies a massive 2.8 trillion parameter Mixture-of-Experts (MoE) architecture. This design allows the model to maintain extreme intelligence while remaining computationally efficient by activating only a subset of parameters for any given task. This efficiency is further bolstered by the new 'Attention Residuals' technique, which has been shown to deliver approximately 25% higher training efficiency at less than 2% additional computational cost.
Perhaps the most significant breakthrough for developers working with massive datasets is the introduction of Kimi Delta Attention. This proprietary mechanism solves the traditional bottleneck of long-context processing. Kimi K3 supports a staggering 1-million-token context window, and thanks to Delta Attention, it achieves up to 6.3x faster decoding speeds in these million-token contexts compared to previous generations. This makes real-time interaction with massive codebases or entire libraries of documents a reality.
- 2.8T Parameter Mixture-of-Experts (MoE)
- Native Multimodal Capabilities (Text, Vision, and more)
- Kimi Delta Attention: Up to 6.3x faster decoding in long contexts
- Attention Residuals: 25% increase in training efficiency
Performance & Benchmarks: Closing the Frontier Gap
Kimi K3 has demonstrated frontier-level performance across a wide array of industry-standard evaluations. In coding benchmarks, the model shows exceptional proficiency in long-horizon tasks, making it a formidable tool for autonomous software engineering. While it currently trails slightly behind the absolute peak of Claude Fable 5 and GPT 5.6 Sol, it significantly outperforms all other currently available open models.