MOONSHOT AI / NATIVE MULTIMODAL / 2026

Kimi K2.5

TOTAL1.06Tparameters

ACTIVE32Bper token

DEPTH61 + ViT27repeated blocks

HIDDEN7,168main trunk

CONTEXT256Knative tokens

CHECKPOINT595.2GB64 shards · Modified MIT

FAMILY EVOLUTION

Kimi LinearKDAKimi K21T MoEK2.5native visionK2.7code agent

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

100K INPUT + 1K OUTPUT · 19.21P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT100.0KIMAGE · VIDEO · TEXT

TRANSFORMER STACK61 TEXT LAYERS + MOONVIT1 dense FFN · 60 sparse MoE · MLA throughout

pre-norm additive residual · media tokens join the same language stream

OUTPUT1.0KTOKENS

wraps every repeated block

DENSE MLA · MOONVITMoonViT 以空间—时间 attention 编码图像与视频，2×2 PatchMerger 将视觉 token 压缩后送入语言主干。

Kimi MoE · 1.06T CAPACITY32B active path · total capacity ≠ active compute

TEXT DEPTH611 dense + 60 MoE

VISION27L≈400M

HIDDEN7,168main trunk

EXPERTS384top-8 + 1

CONTEXT256Ktokens

PRECISIONINT4native QAT