A 25GB Laptop Running a 744B-Parameter Giant?

Plot twist: I thought frontier AI models were data-center property — until colibri. A 744B-parameter giant (GLM-5.2 class), officially runnable on a 25GB-RAM, 12-core laptop. Slow, but it runs. 30k stars — a project turning ‘impossible’ into ‘slow but possible.‘

The Giant Killer for Consumer Hardware

colibri is an inference engine that runs huge AI models on your own computer — no cloud, no data center. Its secret is memory tiering: treat GPU memory, RAM, and disk as one warehouse, streaming the ‘experts’ the model needs from disk. The model doesn’t need to fit in memory — it needs to be placed.

Features

Memory tiering is the counterintuitive core: a 744B model activates only ~40B params per token — 11GB of ‘active experts’ shifting per token. That’s the physics behind ‘thin-hardware running.‘

The Hummingbird Philosophy

colibrì is Italian for ‘hummingbird’ — ‘a few grams, hovering, visiting a thousand flowers; this engine feeds a 744B giant with hummingbird rations.’ Built by one person on a 12-core, 25GB laptop. The belief: frontier models shouldn’t be sealed in data centers — curious people should open, observe, improve them.

llama.cpp or This?

Ask what you want to run: daily small models → Ollama/llama.cpp; ‘models I never dared dream of’ → colibri is the only choice.

Trust Check

Solo dev, little public info — but the project itself is transparent: official community speed tables (M5 Max, DGX Spark, 6×RTX 5090), all numbers public. Apache 2.0, solo-maintained, very new (July 2026) — 30k stars in two months is the community’s verdict on breaking rules.

Where It’s Headed

‘An inference engine you can use today + an open research platform’ — more model families, better compression, lower hardware barriers and cost per token. One sentence: ‘Not renting intelligence — owning it.‘

Install

Tested on macOS 15 (16GB):

git clone https://github.com/JustVugg/colibri.git && cd colibri/c
./setup.sh        # checks gcc/OpenMP, builds, self-tests
./coli doctor     # diagnose missing pieces
./coli chat

Build clean. Gotchas: ① don’t expect speed — 25GB cold start is 0.05-0.1 tok/s; ② model containers ~372GB; ③ slow-but-correct is normal, not a fault.

FAQ

• Q: Can my machine run it? A: 25GB RAM + disk patience starts it — speed depends on hardware.
• Q: Usable speed? A: For personal research; for speed rent GPUs. Its value is ‘can run,’ not ‘runs fast.’
• Q: Commercial? A: Apache 2.0 — check model licenses.

Technologist or pragmatist? Comment which — the hummingbird engine was built for the technologists.