Moonshot AI’s Kimi K3: First Open 3T AI Model Launches

Moonshot AI has unveiled Kimi K3, a 2.8-trillion-parameter open-weight model, marking a major step in the AI arms race. Released on July 27, 2026, Kimi K3 is the first openly available model in the 3-trillion-parameter class, positioning itself as a serious challenger to proprietary models like OpenAI’s GPT 5.6 Sol and Anthropic’s Claude Fable 5.

The model, developed by Beijing-based Moonshot AI, is designed for coding, long-context reasoning, and multimodal tasks. Its standout feature is the 1-million-token context window, the largest among open models, allowing for extensive knowledge and reasoning capacities across applications.

Technical Innovations and Performance

Kimi K3 incorporates several architectural updates aimed at tackling the challenges of scaling large models:

Kimi Delta Attention (KDA): Enables efficient processing of long-context inputs up to 1 million tokens.
Attention Residuals: Improves the model’s depth-wise reasoning by selectively retrieving information.
Stable LatentMoE Framework: Only 16 of 896 experts are activated per token, optimizing performance while keeping compute costs manageable.

Benchmark results show Kimi K3 outpacing proprietary competitors in several areas. Notably, it leads on coding benchmarks like SWE Marathon and BrowseComp while remaining close to GPT 5.6 Sol and Claude Fable 5 on other tasks. According to Moonshot, the model achieves frontier-level performance across agentic and reasoning benchmarks, highlighting its versatility in both structured and creative tasks.

Pricing and Accessibility

With an aggressive pricing structure, Kimi K3 could disrupt the commercial AI model market. API pricing is set at $0.30 per 1 million tokens for cache-hit inputs and $3.00 per 1 million tokens for cache misses, significantly undercutting proprietary offerings. Output tokens are billed at $15.00 per 1 million tokens. This pricing incentivizes developers to optimize for cache efficiency, potentially reducing costs in production environments.

The model is available via the Together AI API, which is compatible with OpenAI’s API. Developers can also download the open weights to run the model independently, an option unavailable with closed-source competitors like GPT-4 or Claude.

Why Kimi K3 Matters

Kimi K3’s release highlights a growing shift in the AI landscape: the race to develop open-source alternatives with capabilities rivaling closed models. By offering transparency and flexibility, Moonshot AI targets developers and enterprises seeking high-performance models without vendor lock-in. Its success could boost adoption of open-weight models across industries, setting a precedent for future large-scale AI development.

The model’s open nature also aligns with increasing scrutiny over the proprietary nature of AI and its implications for control, accessibility, and innovation. Kimi K3’s 3T-class capabilities signal that open AI is catching up with—and in some cases outperforming—closed ecosystems.

What’s Next?

Demand for Kimi K3 has reportedly overwhelmed Moonshot AI’s capacity, leading to a temporary pause on new API subscriptions. This underscores the strong interest in high-performance open models and the growing appetite for alternatives to dominant players like OpenAI and Anthropic.

Moonshot has positioned Kimi K3 as a foundation model for a range of applications, from enterprise software development to advanced research tasks. With its competitive pricing and transparent architecture, Kimi K3 could reshape the economics of large-scale AI deployment. Interested developers can start building today via the Together AI documentation.

Image source: Shutterstock