Breaking

Moonshot AI Unveils Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model Closing the Gap With US Rivals

July 18, 2026 at 00:16 EDT

Foundation Models
Open Source

China's Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open-weight model that its developers say rivals the best systems from OpenAI and Anthropic, reviving fears in the US of a repeat of the "DeepSeek moment" that rattled markets a year earlier. The model, unveiled around July 16-17, 2026, is already available through Moonshot's Kimi app, website and API, with the full model weights due to be published on July 27.

July 16–17, 2026 · Moonshot AI

Kimi K3: China's 2.8-Trillion-Parameter Open Model Closes In on the Frontier

Beijing-based Moonshot unveils the first open-source model in the three-trillion class — a direct challenge to Claude and GPT-5 that has revived fears of another "DeepSeek moment" on Wall Street.

2.8T

Parameters — first open model in the 3-trillion class

Token context window for long-horizon work

16 / 896

Experts activated in the Stable Latent MoE design

Jul 27

Full open weights to be published

Scaling efficiency vs. the earlier K2 system

Moonshot claims K3's new architecture scales roughly 2.5× more efficiently than K2.

1×

2.5×

Where it lands on the leaderboards

Frontend Code Arena — 1,679 points

Vals Index

Near top

Artificial Analysis — behind Claude & GPT-5

API pricing listed around $3 / 1M input tokens · $15 / 1M output tokens — open weights allow local inference.

The "DeepSeek moment," again

Investors treated K3 as another China surprise, fueling a tech selloff. The SOX semiconductor index fell roughly 20% from its recent high — with some analysts calling it partly profit-taking amid overinvestment worries.

Recent high

−20%

SOX index

Developer optimism

Strong agentic coding and stability on long, complex projects
High interest in running weights locally via Hugging Face and Ollama
Seen as a practical substitute for closed models on everyday tasks

Caution

Full evaluation must wait until weights ship on July 27
Open questions on real-world reliability under load
Regulatory scrutiny of a Chinese-built frontier model

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.