Moonshot AI launches Kimi K3, an open‑source model that narrows the AI gap

Moonshot AI launched the Kimi K3, a 2.8‑trillion‑parameter model it claims to be the largest open‑source AI system in the world, with full weights scheduled for release on July 27. Benchmark tables show K3 scoring in the 1–1.7k range on GDPval‑AA v2, close to proprietary models, and achieving 91.2/100 on BrowseComp.

Architecturally, K3 relies on Moonshot’s custom Kimi Delta Attention and Attention Residuals, both open research. It offers a 1‑million‑token context window and a “thinking mode” feature. Though the components are published, their combination at this scale is novel.

Pricing is a key differentiator: $3 per million input tokens, $15 per million output tokens, and $0.30 for cached inputs. While competitive against mid‑tier APIs, running a 2.8‑trillion‑parameter model still demands a disaggregated serving infrastructure (Mooncake الأربعاء).

Two demos highlight potential: a 48‑hour autonomous chip‑design run yielding a 4‑mm² layout at 100 MHz and an I‑Love‑Q astrophysics calculation reduced from weeks to ~2 hours. These serve as proof‑of‑concepts rather than production ready.

For enterprise adoption, the practical questions are: reproducibility under independent evaluation, self‑hosting.hidden costs, and the reliability of agentic capabilities over long sessions. These remain open as the weights go public.

The release signals that open‑source can close the performance gap faster than the “six‑monthmarkt” narrative suggested. K3 joins Moonshot’s K2.7 Code ($0.95/$4) and K2.6 ($0.95/$4) in a three‑tier lineup aimed at developers and enterprises alike.