Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-30 · high
A reverse-engineered parameter count puts Kimi K3's smallest deployable tier near 870GB, past the 512GB ceiling on the biggest Apple Silicon machine sold.
Why now: The repo went up July 27, directly countering single-Mac claims about Kimi K3 that were circulating the same week.
Kimi K3's full architecture doesn't fit on a single Mac, according to a reverse-engineered parameter map published on GitHub July 27. The repo, from PipeNetwork, has pulled in 284 stars for laying out exactly what's inside the model.
That matters for anyone budgeting hardware to run frontier open-weight models locally. Routed experts alone account for 97.94% of Kimi K3's 2,780B total parameters, per the repo's breakdown. That accounting validates exactly against the model's published 1.561TB repo size.
The detail is granular. 896 routed experts get selected at top-16 per token, plus 2 shared experts. They spread across 93 layers split 69 Kimi Delta Attention to 24 gated MLA. A SiTU-GLU activation function and an AttnRes mechanism that mixes the residual stream every 12 layers aren't things I've seen documented elsewhere. The LatentMoE experts run in a compressed 3584-dimension space, half the model's 7168-d residual stream.
The README doesn't hedge on this point. The smallest deployable tier needs roughly 870GB, and the largest Apple Silicon machine on the market tops out at 512GB. No single Mac clears that gap.
Kimi K3 was built as a cluster model, not a local one. Any claim of running the full weights on a single Mac is quantized, distilled, or false. Check the math before you believe one.
The repo went up July 27, directly countering single-Mac claims about Kimi K3 that were circulating the same week.
Each link below shares sources, entities, or timing with this story.
Kimi K3 built by Moonshot / Shared entities / Same source domain / Shared topic / Earlier coverage
Linked by a graph relationship (Kimi K3 built by Moonshot); both cover AttnRes, Kimi Delta Attention, Kimi K3; reported by the same outlet (github.com).
Kimi K3 uses NoPE / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Kimi K3 uses NoPE); both cover July, Kimi Delta Attention, Kimi K3, LatentMoE; overlapping topics (attention, kimi, layer).
Kimi K3 competes with OpenAI / Shared entities / Earlier coverage
Linked by a graph relationship (Kimi K3 competes with OpenAI); both cover July, Kimi Delta Attention, Kimi K3; earlier July coverage from 2026-07-21.
Linked by a graph relationship (Kimi K3 competes with OpenAI); both cover July, Kimi K3; earlier July coverage from 2026-07-20.
Kimi K3 built by Moonshot / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Kimi K3 built by Moonshot); both cover July, Kimi K3; overlapping topics (against, kimi).
Kimi K3 built by Moonshot / Shared entity: July / Same source domain / Earlier coverage
Linked by a graph relationship (Kimi K3 built by Moonshot); both cover July; reported by the same outlet (github.com).
Kimi K3 benchmarked against Fable / Shared entity: July / Earlier coverage / Tension
Linked by a graph relationship (Kimi K3 benchmarked against Fable); both cover July; earlier July coverage from 2026-07-27.
Linked by a graph relationship (Kimi K3 benchmarked against Fable); both cover July; earlier July coverage from 2026-07-24.