Kimi K3: First Open-Source LLM with 2.8 Trillion Parameters

Key Points at a Glance

• Moonshot AI releases Kimi K3, the first open-weight model in the 3-trillion-parameter class.
• The highly efficient MoE architecture offers a context window of 1 million tokens, native image understanding, and activates only 16 out of 896 experts.
• The API has been available since mid-July 2026, with the release of the model weights scheduled for July 27, 2026.

Moonshot AI, the up-and-coming AI company behind the Kimi platform, announced Kimi K3 in July 2026 —the world’s largest open-weight AI model to date, which , with 2.8 trillion parameters, competes directly with proprietary systems from OpenAI and Anthropic. As Moonshot AI states in its official documentation, the flagship model combines massive scale with an extremely large context window and native multimodality.

The innovations in detail

According to Moonshot AI’s technical specifications, Kimi K3 is based on an extremely large yet resource-efficient Mixture-of-Experts (MoE) architecture. The model has a total of 2.8 trillion parameters. However, to keep inference costs and computational overhead manageable, only 16 out of 896 experts are activated per token.

A standout technical feature is the 1 million-token context window. This capacity is supported by new architectural approaches such as Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). In addition, the system offers automatic and free context caching for input lengths of 256 tokens or more.

Unlike pure text models, Kimi K3 has native image and video understanding. For developers, the model also integrates “Always-on Reasoning,” Dynamic Tool Loading, and guaranteed structured JSON output—essential features for autonomous coding tasks and complex agent systems.

Why This Matters

The release of Kimi K3 marks a turning point in AI development: It is the first time a model in the so-called 3T class (3 trillion) has been made available to the general public as an open-weight release. By comparison, open-source competitors such as DeepSeek V4 Pro have around 1.6 trillion parameters, according to industry reports.

With this scale, Moonshot AI is directly challenging the proprietary market leaders. The massive context window, paired with strong reasoning capabilities, makes K3 particularly attractive for complex agent workflows and the analysis of massive codebases. The reliability of the models is crucial, especially for in-depth reasoning processes. As recent analyses show, even among market leaders like OpenAI, reasoning models can hallucinate more frequently if the parameters aren’t optimally tuned for long logical chains. Kimi K3 aims to solve this problem through its specialized KDA architecture.

Kimi K3 delivers frontier-class performance to the open-source community and is forcing proprietary providers to rethink their pricing and access models.

However, this is also the main point of criticism from the developer community: Kimi K3 is an open-weight model, not a true open-source project in the strict academic sense. The weights are shared under a modified MIT license, but the training data and pipeline remain proprietary. Furthermore, a model with 2.8 trillion parameters requires massive hardware resources for hosting. To alleviate this bottleneck, ecosystem tools such as vLLM have already announced “Day-0” support to facilitate deployment on enterprise clusters.

Availability & Conclusion

Access to Kimi K3 is initially available via the official API, which launched in mid-July 2026. According to the developers, the full model weights are scheduled to be released on July 27, 2026. Conclusion: Kimi K3 drastically pushes the boundaries of what’s possible outside of closed AI labs, but due to its enormous hardware requirements for self-hosting, it’s primarily aimed at enterprise users and professional development teams.

FAQ

What is Kimi K3?
Kimi K3 is an AI language model from Moonshot AI with 2.8 trillion parameters. It is currently the world’s largest open-weight model, offers a context window of 1 million tokens, and features native image and video understanding.

When will Kimi K3 be available?
The API for Kimi K3 has been available since mid-July 2026. The release of the open model weights is scheduled for July 27, 2026, allowing developers to host the model on their own hardware.

Is Kimi K3 truly open source?
No, it is an open-weights release. This means that the final model weights are freely available under a modified MIT license. However, Moonshot AI keeps the underlying training data and training pipeline under wraps.

How does Kimi K3 differ from DeepSeek?
The main difference lies in scalability and architecture. While current open DeepSeek models such as V4 Pro have around 1.6 trillion parameters, Kimi K3, with 2.8 trillion parameters, is the first open model to break into the 3T class.