Breaking

Moonshot AI Launches Kimi K3, a 2.8-Trillion-Parameter Multimodal Model, With Open Weights Due July 27

July 17, 2026 at 09:45 EDT

Foundation Models
Open Source
Software Dev & Coding

Moonshot AI on July 16, 2026 released Kimi K3, a natively multimodal model with 2.8 trillion parameters and a 1-million-token context window, positioning the Beijing-based company among the world's most ambitious open-weight developers. The model is available now through APIs including OpenRouter and Moonshot's own platform, with the full weights scheduled for public release on July 27.

July 16, 2026 · Moonshot AI

Kimi K3: A 2.8-Trillion-Parameter Open Challenger

Beijing's Moonshot AI launches a natively multimodal model with a 1-million-token context window — the first open-source system at the three-trillion-parameter class, taking direct aim at frontier models from OpenAI and Anthropic. Full weights arrive July 27.

2.8T

total parameters — largest open-weight class

token context window

16/896

experts active per pass (~1.8%)

2.5×

scaling efficiency vs. K2

Where K3 stands on the leaderboards

Rank across three independent evaluations — top tier, not yet outright leader.

Frontend Code Arena
1679 Elo

Vals Index

AA Intelligence Index
57.11 (behind Claude Fable 5, GPT-5.6 Sol)

Trails leading closed systems on aggregate intelligence — but matches or beats them on coding & interface tasks.

Developer praise

Polished UI & design output
Strong frontend, 3D & game generation
Fast responses relative to model size
Reliable on long-running agentic tasks

Caveats

Can expand a task beyond what was asked
Burns tokens faster than K2.6
Pricing near Anthropic's range: $3 in / $15 out per M tokens
No local verification until weights ship July 27

The bigger picture

A rapid cadence — K2 → K2.5 → K2.6 → K2.7 Code → K3 in one year — shows Chinese labs closing the gap despite compute constraints.

If the open weights match the API results on July 27, K3 could become a widely tested reference point for frontier-scale open models — intensifying competition over long-context, agentic and multimodal capabilities.

Continue reading

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