Breaking

Moonshot AI Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model Rivaling Top U.S. Systems

July 18, 2026 at 07:08 EDT

Foundation Models
Open Source
Software Dev & Coding

China's Moonshot AI on Thursday released Kimi K3, a 2.8-trillion-parameter open-weight model that matches or beats leading American systems on several benchmarks and jolted markets in what analysts are calling a second "DeepSeek moment." The model, from the Beijing-based startup, went live immediately on Moonshot's API and Kimi platform, with the full model weights scheduled for public release on July 27.

July 16, 2026 · Moonshot AI

Kimi K3: China's 2.8-Trillion-Parameter Open Model Triggers a Second "DeepSeek Moment"

Beijing-based Moonshot AI released a natively multimodal, open-weight model that matches or beats leading U.S. systems on several benchmarks — jolting markets and narrowing the frontier gap under U.S. export controls.

2.8T

parameters — largest open-source model yet released

token context window for long-horizon tasks

16/896

experts activated per token (Mixture-of-Experts)

blind human-preference leaderboard for frontend coding

Cache-Hit Pricing Collapses the Input Cost

Coding workloads routinely hit >90% cache-hit rates — cutting input cost by 10× vs a cache miss. Cost per 1M input tokens, drawn to scale.

$3.00

cache miss

$0.30

cache hit

Output: $15 / 1M tokens. Mid-tier pricing — not the rock-bottom levels of some earlier Chinese releases, but effective cost falls sharply with high cache-hit rates.

How Kimi K3 Stacks Up

Competitive with

Anthropic Fable 5

Surpasses

Anthropic Opus 4.8 · beats GPT-5.6 Sol on some tasks

Leads

Frontend coding · strong Terminal-Bench · top-3 GDPval-AA v2

Enthusiasm

Praised by developers as the best open-source model available — practical for coding and agentic work, with calls to integrate it into tools like Cursor. Open weights (July 27) enable local run and fine-tuning.

Caution

Real-world reliability untested so soon after launch. Open questions on whether benchmarks overstate capability, and whether gains come from distillation or genuine architectural advances.

Market Response Was Swift

Nvidia ▼

shares fell

TSMC ▼

shares fell

SoftBank ▼

shares fell

Built on domestic hardware including Huawei chips, the launch has investors weighing

Continue reading

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