Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-16 · high
At roughly one token a second, it's built for overnight batch jobs, not chat, on hardware you already own.
Why now: Two weeks after its July 1 launch, Colibri's star count is still climbing.
Colibri streams GLM-5.2's 744 billion parameters off disk to run the model on a 25GB consumer machine, per its GitHub repo.
That swaps a GPU cluster for a single desktop most engineers already own, at the cost of speed. The model's 370GB of routed experts have to be pulled from disk for nearly every token, since VRAM and RAM can't hold them.
Colibri, built by developer JustVugg as a dependency-free C engine, splits the model in two. About 17B dense parameters stay resident in memory as int4, roughly 9.9GB. The other 19,456 routed experts, spread across 75 MoE layers, sit on disk instead, at about 19MB each.
The benchmarks don't dress up the cost. A Ryzen 9 9950X with Gen5 NVMe hits 1.23 tokens per second. An M5 Max manages 1.06.
A 25GB WSL2 box with a slower drive drops to 0.05 to 0.1 tokens per second. Cold decode alone pulls about 11GB off disk for a single token.
The repo doesn't say how hardware between those extremes would perform.
Native MTP speculative decoding claws some of that back, producing 2.2 to 2.8 tokens per forward pass, but it doesn't change the basic trade.
At roughly one token a second, nobody's chatting with this in real time. What Colibri actually enables is running a 744B model as an overnight batch job on hardware you already own, no cloud bill required.
Since its July 1 release, the project has picked up 14,728 stars and 1,273 forks on GitHub, about 981 stars a day.
Each link below shares sources, entities, or timing with this story.
Shared entities / Same source / Shared topic / Earlier coverage
Both cover Colibri, JustVugg, MoE, RAM; cite the same source (JustVugg/colibri); overlapping topics (colibri, expert, token).
M5 Max benchmarked against M1 Max / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (M5 Max benchmarked against M1 Max); both cover July, MoE, VRAM; reported by the same outlet (github.com).
Shared entities / Same source domain / Shared topic / What happened next
Both cover MoE, NVMe, RAM; reported by the same outlet (github.com); overlapping topics (cost, dense, expert, nvme, token).
Both cover July, NVMe, RAM; reported by the same outlet (github.com); overlapping topics (cold, expert, nvme, token).
Lumabri supports GLM / Shared entities / Same source domain / What happened next
Linked by a graph relationship (Lumabri supports GLM); both cover GLM, JustVugg, MoE; reported by the same outlet (github.com).
M5 Max benchmarked against M1 Max / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (M5 Max benchmarked against M1 Max); both cover July, NVMe; reported by the same outlet (github.com).
Shared entities / Same source domain / What happened next
Both cover GLM, July, MoE, NVMe; reported by the same outlet (github.com); picks up the GLM thread on 2026-08-07.
M5 Max benchmarked against M3 Max / Shared entities / Same source domain / What happened next
Linked by a graph relationship (M5 Max benchmarked against M3 Max); both cover July, RAM; reported by the same outlet (github.com).