Laguna · Inkling · Kimi K3

primary sources only · not invented parity


axisLaguna S 2.1InklingKimi K3
total/active118B/8B~975B/41B2.8T/?
MoE10/256+1 add6/256+2 sink16/896
routersigmoidsigmoid+sinkLatentMoE
layers48 · 12G/36SWA66 · 11G/55SWATBD
PE / long-ctxYaRN+SWA RoPErelativeKDA
attn scale1/√d + softplus gate1/d + SConvTBD
mmtextimage+audiovision+video
weightspublic OpenMDWpublic ApacheAPI-first
TB2.1 (Poolside table)70.263.888.3*
focusagentic codingcustom generalistopen frontier product

Prefer Laguna for mid-size open MoE + day-0 serve stack. Prefer Inkling for relative PE + audio + transformers path. Prefer Kimi K3 for scale + product API / KDA theory.

* peer TB scores from Poolside tables; harnesses differ. gaps: K3 active · K3 layer table · matched bake-offs

← home