Kimi K3: Inside the Largest Open-Source AI — and How It Compares to OpenAI & Anthropic

Overview

Kimi K3 is the largest open-source AI model built to date — a 2.8-trillion-parameter Mixture-of-Experts from Moonshot AI, released in July 2026 with its full weights published for anyone to download, run, or fine-tune. It scores about 57 on the Artificial Analysis Intelligence Index (roughly 3rd overall), essentially matching Anthropic's Opus 4.8 and OpenAI's GPT-5.5 while undercutting both on price — and it takes the #1 spot on front-end coding.

In this video

• Kimi K3 is the largest open-source AI model released so far, with 2.8 trillion total parameters.
• It comes from Moonshot AI, a Chinese lab, and shipped in July 2026.
• It uses a Mixture-of-Experts design — 896 expert networks, but only about 16 activate per token, so inference stays efficient despite the massive size.
• A new attention mechanism, Kimi Delta Attention, lets it handle a 1-million-token context window.
• Reasoning is on by default: every query gets deliberate, step-by-step thinking rather than a quick answer.
• "Open weights" means Moonshot publishes the trained model's parameters publicly, so anyone can download, run, fine-tune, or self-host it — the full weights land July 27.
• By contrast, OpenAI's GPT and Anthropic's Claude are closed models: their weights are proprietary and reachable only through a paid API.
• On the Artificial Analysis Intelligence Index, Kimi K3 scores about 57 — roughly 3rd overall — matching Opus 4.8 and GPT-5.5, and trailing Claude Fable 5 and GPT-5.6 Sol.
• On front-end coding it ranks #1, ahead of Claude Fable 5, and it undercuts both OpenAI and Anthropic on cost.

Frequently asked questions

What is Kimi K3?

Kimi K3 is a 2.8-trillion-parameter, open-weights Mixture-of-Experts AI model from Moonshot AI, released in July 2026 — the largest open-source model built so far.

What does open source (open weights) mean for an AI model?

It means the trained model's weights are published publicly, so anyone can download, run, fine-tune, or self-host the model themselves — unlike closed models that are only reachable through a paid API.

How does Kimi K3 compare to OpenAI and Anthropic's top models?

It scores about 57 on the Artificial Analysis Intelligence Index — roughly 3rd overall — matching Opus 4.8 and GPT-5.5, and trailing Claude Fable 5 and GPT-5.6 Sol, while costing less than both.

How can a 2.8-trillion-parameter model run efficiently?

It is a Mixture-of-Experts: of its 896 expert networks, only about 16 activate for any given token, so the compute per token stays far lower than the total parameter count suggests.

Is Kimi K3 better than Claude or GPT at anything?

Yes — it takes the #1 spot on front-end coding, ahead of Claude Fable 5, and beats both GPT-5.5 and Opus 4.8 on some agentic tasks, while undercutting them on price.