MOONSHOT AI / 2026

Kimi K3

A native multimodal 2.8T-class model for long-horizon coding, knowledge work, and million-token contexts.

Open in workbench Static fit only Open primary source

TOTAL PARAMETERS2.8T

ACTIVE / TOKEN104B

PINNED ARTIFACT1.56 TB

MAX CONTEXT1M

DECISION BRIEF

Long-horizon agents

Stable LatentMoE routes each token through 16 of 896 experts. Its native artifact is still 1.56 TB, making exact residency and expert communication decisive deployment constraints.

WEIGHTSOpen weights

Publicly available checkpoints; verify terms before commercial deployment.

LICENSEUnverified in registry

Unknown remains explicit until the exact checkpoint license is verified and sourced.

EVIDENCEPrimary source

Architecture values are linked to the publishing organization.

ARCHITECTURE

Compute is sparse. Residency is not.

Stable LatentMoE routes each token through 16 of 896 experts. Its native artifact is still 1.56 TB, making exact residency and expert communication decisive deployment constraints.

Expert topology: 896 / top-16. The active-parameter count approximates token-level compute; it does not determine checkpoint memory, KV-cache demand, expert placement, or interconnect pressure.

PRIMARY SOURCE

Moonshot AI — Kimi K3

Inspect evidence

PINNED ARTIFACT

moonshotai/Kimi-K3@9f62e4e9fffb · 1.56 TB

Open manifest

NEXT ACTION

Test the workload, not the headline.

Start with checkpoint residency, then layer in context, concurrency, runtime overhead, topology, and resilience.

Build an assurance plan