China's Moonshot AI abruptly released Kimi K3, a 2.8-trillion-parameter flagship model that rivals top U.S. systems, sending a jolt through markets and forcing investors to reconsider core assumptions behind the global technology rally. The Beijing-based startup made the model available via its API on July 16, 2026, with full open weights scheduled to follow on July 27.
July 16, 2026 · Moonshot AI
Kimi K3 Arrives: China's 2.8-Trillion-Parameter Challenger to the AI Frontier
A near-silent launch put an open-source flagship rivaling top U.S. systems into the market overnight — undercutting premium pricing, jolting equities, and pressuring semiconductor shares. Full open weights follow July 27.
2.8T
total parameters — one of the largest open-source models ever
896
MoE experts, only 16 active per token for efficient inference
1M
token context window with always-on reasoning & native vision
2.5×
scaling-efficiency gain claimed over the prior K2 generation
GDPval-AA v2 — where K3 lands on the frontier
Higher score = stronger. 44 professions across 9 industries.
1,815
Claude Fable 5 Max
1st
1,748
GPT-5.6 Sol Max
2nd
Also 4th of 189 on the Artificial Analysis Intelligence Index (57.1) and 1st on the Arena.ai Frontend Code leaderboard (1,679).
Output pricing vs top U.S. tier
Per million output tokens — K3 is China's priciest model, yet still about half the cost of the most capable U.S. offering.
Input: $3 /M (cache miss) · $0.30 /M (cache hit) — flat across the full 1M-token window.
What impresses
Near top-tier coding and vision performance
Long context aids complex code & agentic workflows
Open weights let teams self-host (July 27)
The caveats
Verbose output raises token use & effective cost
Self-hosting still needs high-end GPUs
Reasoning efficiency seen as on par or slightly behind
The market question
Semiconductor shares fell as investors reassessed bets on U.S. dominance and premium pricing — but analysts flag a possible reversal.
Cheaper per-token inference
→
More overall AI usage
→
Rising demand for memory & storage
A potential Jevons-paradox effect — the coming weeks test whether K3 durably narrows the frontier gap or proves hard to run at scale.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…