Kimi K3 Arrives: Moonshot AI Unveils the World’s Largest Open Model

by Zaira Varghese 18. July 2026 AI 0

Kimi K3 has arrived, and Chinese lab Moonshot AI is calling it the largest open-weight model ever released — a 2.8-trillion-parameter system that early benchmarks show trading blows with the best proprietary AI from OpenAI and Anthropic.

Unveiled on July 16, the model lands at a moment when the distance between America’s frontier labs and China’s open-source challengers looks thinner than at any point in the AI race. Moonshot is not just matching Western systems on paper; on at least one leaderboard, it is beating them outright.

What makes Kimi K3 different

At 2.8 trillion parameters, Kimi K3 is a mixture-of-experts model with a one-million-token context window, native visual understanding, and an always-on reasoning mode Moonshot calls “thinking mode.” Under the hood sit two homegrown architectural tricks — Kimi Delta Attention, a hybrid linear attention mechanism, and a technique the company calls Attention Residuals — designed to keep a model this large fast and affordable to run.

Inside the benchmark results

The numbers are what turned heads. The model took the number-one spot on the Frontend Code Arena with 1,679 points, edging past Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), and GLM-5.2 (1,587). On the broader Artificial Analysis Intelligence Index it scores 57, ranking fourth out of 189 models — level with Claude Opus 4.8 and GPT-5.5, and behind only Claude Fable 5 and GPT-5.6 Sol. VentureBeat’s breakdown has the full benchmark tables.

Open weights and the China factor

Moonshot says it will release the open weights by July 27, letting anyone download, fine-tune, and self-host the model. That matters both commercially and geopolitically. Every time a Chinese lab ships a competitive open model, it chips away at the assumption that the most capable AI will stay locked inside a handful of US companies — and it hands developers worldwide a free alternative to metered API pricing.

Size: 2.8 trillion parameters, the largest open model to date.
Context: a one-million-token window with native vision.
Ranking: fourth on the Artificial Analysis Intelligence Index, first on Frontend Code Arena.

Why it matters for the AI race

For years the open-source frontier trailed the proprietary one by six to twelve months. Kimi K3 compresses that gap to weeks. If the open weights live up to the benchmarks when they land on July 27, Moonshot will have handed the global developer community a near-state-of-the-art model for free — and put fresh pressure on OpenAI, Anthropic, and a stumbling Google to justify their premiums.

The catch, as always, is independent verification. Leaderboard scores are not the same as real-world reliability, and open-weight releases often behave differently once the community stress-tests them. But even if Kimi K3 only lands close to its billing, it is a milestone: the clearest sign yet that open models are no longer content to follow.