Moonshot AI Releases 2.8-Trillion-Parameter Kimi K3 Model

On July 16, 2026, Beijing-based AI startup Moonshot AI unveiled Kimi K3, its latest large language model featuring 2.8 trillion total parameters. Backed by Chinese tech giants Alibaba and Tencent, the company claims the system is now the largest open-source AI model in the world.

The model relies on Mixture-of-Experts (MoE) architecture, activating 16 out of 896 experts to manage compute resources efficiently. It supports a 1-million-token context window and possesses native visual understanding capabilities to parse image files.

The visual below illustrates the scale and release of the Kimi K3 system, presenting the logo on a mobile phone interface.

Users should note the background typography showing "KIMI" alongside the network node visual, which symbolizes the model's new architecture designed to improve how information flows across long sequences. The graphical infinity-loop representation highlights the scaling push from Moonshot AI to challenge Western systems.

Moonshot AI has announced that it will release the full weights of the model to the public on July 27, 2026. In the interim, developers can access the system through the Kimi website or its API, which is compatible with standard OpenAI SDK structures.

Moonshot AI prices the API at three dollars per million input tokens and fifteen dollars per million output tokens, which makes it competitive with Western offerings. If users hit cache memory, the pricing drops to thirty cents per million input tokens.

Initial tests from third-party evaluators show strong results in coding, knowledge tasks, and logical reasoning. On the front-end development leaderboard hosted by Arena.ai, Kimi K3 ranked first, placing it above major US proprietary software systems.

Following the release, Arena Chief Executive Anastasios Angelopoulos expressed strong confidence in the system on the social platform X:

"This may be the single biggest release of the year, and marks the moment that [open-source software] Chinese models have surpassed U.S. models."

On the coding benchmark Code Arena, the system beat Anthropic's recently released Claude Fable 5 model. This achievement raises questions about the common belief that US artificial intelligence companies maintain a lead of several months over Chinese competitors.

Sriram Krishnan, former senior White House policy adviser on artificial intelligence, shared his perspective on X:

"Kimi k3 is a big moment with multiple implications for the entire industry,"

In its technical blog, Moonshot AI acknowledged that its model still lags behind the very best proprietary software in some areas.

"While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Kimi K3 demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models."

The model uses new technical features called Kimi Delta Attention and Attention Residuals. Moonshot AI reports that these upgrades yield an approximate two-point-five times improvement in overall scaling efficiency compared to its previous Kimi K2 model.

Independent testing by the firm Artificial Analysis also placed Kimi K3 immediately behind leading proprietary options on its intelligence index.

The release comes amid growing national security debates in Washington over Chinese tech development, export controls, and "distillation"—the practice of training smaller models on outputs from larger, proprietary systems. Anthropic previously accused Moonshot AI and other Chinese companies of violating service rules by extracting Claude's capabilities.

As Chinese competitors like Zhipu AI and MiniMax release increasingly capable systems, the gap between US and Chinese artificial intelligence is narrowing. The upcoming public weights release on July 27 will likely allow global developers to host and run this 2.8-trillion-parameter system locally, avoiding costly cloud contracts.