Breaking

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

July 18, 2026 at 21:15 EDT

Foundation Models
Open Source
Software Dev & Coding

China's Moonshot AI has released Kimi K3, a 2.8-trillion-parameter model it bills as the largest open-weight system ever built, positioning it directly against leading closed models from Anthropic and OpenAI. The Beijing-based startup made the flagship available across Kimi.com, Kimi Work, Kimi Code and the Kimi API around July 16, 2026, with full model weights slated to follow by July 27.

July 16, 2026 · Moonshot AI

Kimi K3: the largest open-weight model ever built

China's Moonshot AI unveils a 2.8-trillion-parameter Mixture-of-Experts system, taking aim at leading closed models from Anthropic and OpenAI with open weights and aggressive pricing.

2.8T

total parameters — 16 of 896 experts active per token

token context window, with native visual understanding

Frontend Code Arena, edging out Claude Fable 5

Scaling efficiency vs the K2 series

Architectural changes (KDA attention, AttnRes, Stable LatentMoE) deliver ~2.5× the scaling efficiency.

1×

2.5×

Kimi K3

API pricing (per 1M tokens)

input

$15

output

A fraction of comparable Western offerings; cached input priced far lower.

Real-world bug-fix test

2.3× lower cost

Competitive with Fable 5 on the Cline repo — used more tokens and ran longer, spending them on reasoning and verification.

Where it lands on broad intelligence

GPT-5.6 Sol

Claude Fable 5

Kimi K3

Third on broad indices, yet surpassing systems such as Opus 4.8 in some evaluations — with particular strength on long-horizon coding and agentic tasks.

Developer praise

High capability at low cost
Open weights enable local & custom deployment
Strong handling of large codebases and extended tasks

Caveats noted

Raw speed can lag Fable in certain scenarios
Token consumption runs higher
Durable market share still unproven

Full model weights are slated to release by July 27, 2026 — expected to widen local and custom deployments and test whether cost and openness translate into durable market share against U.S. incumbents.

Continue reading

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