Kimi K3: World's Largest Open-Source AI Model Released by Moonshot AI

Chinese AI company Moonshot AI has released Kimi K3, a new artificial intelligence model with 2.8 trillion parameters. It is now the largest open-source AI model ever created.

The model was announced on July 16, 2026. Moonshot AI plans to release the full model weights to the public by July 27, allowing anyone to download and use it.

What Makes It Special

Kimi K3 can process up to one million tokens of text at once — roughly the equivalent of several full-length novels. It can also understand images natively, without needing a separate vision system.

Despite its massive size, the model only uses a small fraction of its brain for any given task. It selects 16 specialist modules out of 896 available ones for each piece of text it processes. This makes it much cheaper to run than its size would suggest.

How It Compares to ChatGPT and Claude

In standardized testing across 44 professional occupations, Kimi K3 scored 1,687 points. That places it third overall — behind Anthropic's Claude Fable 5 (1,815) and OpenAI's GPT-5.6 Sol (1,747), but ahead of Claude's Opus 4.8 (1,600).

In coding tasks specifically, K3 outperformed all competitors on two major benchmarks: SWE Marathon and Program Bench.

Pricing

Using Kimi K3 through Moonshot's API costs $15 per million output tokens. For context, OpenAI's GPT-5.6 Sol charges $60 per million output tokens and Anthropic's Fable 5 charges $75 — making K3 roughly four to five times cheaper.

For repeated tasks, the cost drops even further: cached inputs cost only $0.30 per million tokens, and Moonshot says 90% of coding-related requests hit the cache.

Why Open-Source Matters

Until now, the most powerful AI models have been locked behind corporate APIs. Users send their data to company servers and pay per use, with no ability to inspect or modify the model.

Open-source models change this equation. When Moonshot releases K3's weights, organizations can run the model on their own servers, keeping sensitive data in-house. Researchers can study how it works, and developers can customize it for specific applications.

The hardware requirement is significant — you need roughly 1.4 terabytes of GPU memory to load the model — but it puts frontier AI within reach of universities and large companies that previously had to rely on API access.

What This Means

The release of Kimi K3 signals that the gap between open-source and proprietary AI is closing rapidly. A year ago, open-source models trailed the best proprietary systems by a wide margin. Today, K3 sits in the top three alongside the most advanced models from OpenAI and Anthropic.

For a detailed breakdown of Kimi K3's capabilities and what it means for the AI industry, read the full analysis on Medium.

For more on Kimi K3, visit wan27.org's comprehensive coverage.