Moonshot AI Unveils Kimi-K3: The World's First Open 3T-Class Model

Moonshot AI is launching Kimi-K3, the first open 3T-class model, on July 27, 2026. This next-generation model features a novel architecture and native agentic capabilities for advanced reasoning and coding. It will be released with open weights and an extended context window for large-scale repository analysis.

Key Points

• Kimi-K3 is the world's first open 3T-class model designed for frontier intelligence.
• The model features a new architecture utilizing Kimi Delta Attention and Attention Residuals.
• It includes native agentic capabilities like tool calling, browsing, and multi-step planning.
• An extended context window is provided to support repository-scale code understanding.
• Moonshot AI will release the model weights openly on July 27, 2026.

Sentiment

Technically curious and cautiously optimistic; the community is excited about the 'open' nature of the model but realistic about the staggering hardware and energy requirements.

In Agreement

• The release of open weights for a 3T-class model is a landmark event for transparency in the AI industry.
• Open models are essential for industries requiring absolute data sovereignty or air-gapped environments where cloud APIs are prohibited.
• The model's native MXFP4 format and MoE architecture are crucial for making such a large model even remotely hostable.
• This release provides a rare 'conflict-free' data point to estimate the marginal cost of serving frontier-level models.

Opposed

• Local CPU-based inference for a 3T model is likely too slow to be useful, potentially taking hours or days for a single complex response.
• The privacy argument for local hosting is often overstated, as enterprise cloud providers (AWS Bedrock, Azure) already offer high-level compliance like HIPAA and government-grade security.
• Electricity costs for running massive, inefficient local servers could significantly exceed the cost of using optimized cloud APIs.
• Estimating the costs of closed models (like GPT-4 or Claude) based on K3 is speculative because their actual parameter counts and architectures remain unknown.