Breaking

Moonshot AI Unveils 2.8-Trillion-Parameter Kimi K3, Rattling Global Tech and Chip Stocks

July 17, 2026 at 17:52 EDT

Foundation Models
Open Source
Funding & M&A

China's Moonshot AI released Kimi K3 on Friday, a 2.8-trillion-parameter open-weight model that the company says matches or beats most leading U.S. systems, triggering a sharp selloff across AI and semiconductor stocks. The launch narrowed the perceived gap between Chinese and American frontier models and revived comparisons to the market shock that followed DeepSeek's debut. Bloomberg and Reuters reported the release, with Reuters describing it as the world's largest open AI model.

July 17, 2026 · Moonshot AI

China's Kimi K3 closes the gap — and rattles the market

A 2.8-trillion-parameter open-weight model that Moonshot says matches or beats most leading U.S. systems — triggering a sharp selloff across AI and semiconductor stocks.

2.8T

parameters — world's largest open AI model

token context window, native multimodal

896 / 16

experts total / active per token (Stable LatentMoE)

The market reaction was immediate

Investors read the launch as fresh proof that a well-funded Chinese lab can approach costly proprietary models — for free.

−6%

Asian
semiconductor gauge

−28%

Shares of Chinese
rival Z.AI

Independent benchmark scores

Terminal-Bench 2.188.3%

GPQA Diamond93.5%

AA-LCR74.7%

DeepSWE67.5%

HLE44.3%

Composite indices: Intelligence 57.1 · Coding 76.2 · Agentic 50.1.

The upside

Early testing favorable on coding and agentic work — 3rd of the field at 62/80. Earlier K2.6 users cited up to an 88% cut in coding costs.

The catch

Weights not open until ~July 27, and local deployment is expected to need 64+ GPUs — out of reach for most users.

Why it matters

Moonshot — founded 2023, valued above $20B and now seeking funding near $30B — keeps the U.S.–China AI race and the open-weight debate at full boil. Whether K3 endures in production, or repeats a headline-driven selloff that later stabilizes, hinges on independent replication and the compute enterprises must pay to run 2.8T parameters themselves.

Continue reading

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