Moonshot AI, the Beijing-based lab behind the Kimi assistant, released its new flagship open-source model on July 16, 2026: the Kimi K3. With approximately 2.8 trillion parameters in a mixture-of-experts (MoE) architecture, VentureBeat describes it as the largest open-source model ever released. Its 1 million token context window allows processing entire code repositories or long documents in a single pass.

According to kimi3.online, Kimi K3 jumped from 18th to 1st place on the Frontend Code Arena with a score of 1679, surpassing Anthropic’s Claude Fable 5 — a model considered unbeatable for front-end tasks. This feat is even more significant as it is an open-weight model competing on equal footing with closed systems from OpenAI and Anthropic.

The model focuses on long-horizon coding, agentic workflows, and end-to-end knowledge work — from research to drafting and presentation. The Kimi app includes features like Swarm (parallel task execution) and Goal (autonomous multi-step objectives).

Technical Specifications

  • Developer: Moonshot AI (Kimi)
  • Release: July 16, 2026
  • Parameters: ~2.8 trillion (MoE)
  • Context window: 1,000,000 tokens
  • Access: web (kimi.com), API (platform.kimi.ai), open weights for self-hosting
  • Predecessors: Kimi K2.6 (general flagship) and K2.7 Code (code specialist)

Benchmarks and Performance

The standout result is the Frontend Code Arena, a blind evaluation where humans vote on which model generates the best front-end code. Kimi K3 leads in six of seven monitored categories, surpassing Claude Fable 5. TechCrunch noted that the release aims to close the gap with Anthropic’s Opus 4.8 on broader benchmarks. Early tests indicate the model also produces playable multiplayer and 3D games from a single prompt.

Comparison with Competitors

Compared to Kimi K2.6, the K3 represents a generational leap: from #18 to #1 on Frontend Code Arena, with a quadrupled context window (from 256K to 1M tokens) and native agentic capabilities. Against Claude Fable 5, the K3 wins in front-end voting and costs significantly less via API ($3.00 per million input tokens vs. Anthropic’s premium pricing), plus offers open weights. GPT-5.x maintains an advantage in ecosystem and multimodality, but the K3 delivers frontier performance at open-source cost.

How to Use

Kimi K3 can be used for free on kimi.com with daily limits. For heavy use, there are paid plans with higher capacity and parallel Swarm execution. The API is compatible with OpenAI clients: simply point the base URL to Moonshot’s endpoint and use the model identifier. Open weights can be downloaded for self-hosting, but require robust clusters — quantized versions and inference providers should emerge quickly.

Availability and License

Moonshot publishes the weights openly, which the press calls ‘the largest open-source model ever released.’ It is recommended to check the license file in the official repository before using in production.

Kimi K3 reaffirms Moonshot AI’s commitment to openness and democratization of artificial intelligence. For teams needing long-context coding and agentic pipelines at controlled cost, the K3 is an option not to be ignored.