China's Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open-weight model that matches or beats leading US systems on several benchmarks, triggering a slide in AI and semiconductor stocks and prompting fresh debate over how far the country trails American labs.
July 17 · Moonshot AI · World AI Conference, Shanghai
China's Kimi K3 cracks the frontier — the largest open-weight model ever built
A 2.8-trillion-parameter open-source model from Beijing matches or beats leading US systems on several benchmarks, undercuts them on price, and reignites the debate over how far China really trails American labs.
2.8T
parameters — the largest open-source model ever released
1.8%
of experts active per token (~16 of 896) via Mixture-of-Experts
1M
token context window, with native vision & multimodal support
$20B
Moonshot valuation; ARR reported above $200M
Intelligence Index — near the top, just short of the best
Artificial Analysis Intelligence Index score (higher is better)
Where K3 lands on independent leaderboards
1st
Frontend Code Arena
1,679
2nd
AA-Briefcase (agentic)
1,527
3rd
GDPval-AA v2 (professional tasks)
1,687
The pitch: frontier performance, ~40% cheaper
Flat rates across the full context, undercutting comparable frontier models by a wide margin.
$0.30
per 1M input tokens (cache hit)
$3
per 1M input tokens (cache miss)
What impresses
Best-in-class price-to-performance, especially frontend coding
Open weights (July 27) enable local & customized deployment
Demand so high sign-ups were paused within 48 hours
The caveats
Overall capability still sits marginally behind the very best US systems
Weight release invites closer scrutiny of true frontier reach
Launch triggered a slide in AI and semiconductor stocks
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…