A big model fits if most of it stays on disk.

Frontier MoE on hardware you already own sounds impossible until you read the last clause. Mixture of experts activates a fraction of the network per token, so the rest can sit on disk and be paged in. The claim rests entirely on that.

make

github.com/JustVugg/colibri

Why streaming is possible at allIllustrative, no real counts

active for this token

paging in from disk

resident on disk, idle

what makes the claim workonly a fraction runs per token no expert count published

pure CNo runtime to install and nothing to resolve. On a machine you already own, the install is the first place a project like this fails.

zero depsSame argument stated twice, which suggests the author has watched dependency trees defeat local inference before.

experts streamed from diskThe mechanism. It moves the bottleneck from memory capacity to disk read speed, which the description does not mention.

The bottleneck moves, it does not vanish

Paging experts per token turns a memory problem into a read-latency problem. Whether that trade works depends on the disk, and no disk, throughput figure, or token rate appears in the description.

What is not on this page

No model, no expert count, no memory floor, no tokens per second. The eight by four grid is a drawing, and none of its cells stand for a real expert.

Concept mock, not affiliated with the project. The complete GitHub Trending description on 10 September 2026 is "Run frontier MoE models on hardware you already own, pure C, zero deps, experts streamed from disk." The 32 squares are an illustrative drawing of sparse activation, not a real architecture: no expert count, routing rule, model, or measurement appears here, because that sentence publishes none.