A pure-C inference engine called Colibrì has demonstrated that Z.ai's 744-billion-parameter GLM-5.2 can run on an ordinary consumer machine with just 25GB of RAM and no GPU, challenging the assumption that frontier-scale models require frontier-scale hardware.
June 2026 · Colibrì × GLM-5.2
A 744-billion-parameter model, running on 25GB of RAM — no GPU
Colibrì, a proof-of-concept engine written in ~2,400 lines of pure C, streams GLM-5.2's sparse experts from disk — proving frontier-scale open weights can run on an ordinary consumer machine. The catch: speed.
744B
total parameters — but only ~40B active per token
25GB
RAM footprint — instead of a 1.5TB-class load
0 GPU
pure C, zero dependencies, optional CUDA
How 25GB holds a 1.5TB-class model
The dense network stays in RAM; the routed experts live on disk and stream on demand. Only a fraction activates per token.
~9.9GB
Dense layers
~370GB
Int4 experts
The trade-off is speed
Cold generation is bottlenecked by disk I/O per token. Better NVMe and more RAM could close the gap — in theory.
0.05–0.1
tok/s cold
→
5–15
tok/s target
Where GLM-5.2 lands on coding & agentic tasks
SWE-bench Pro — GLM-5.2 62.1
Terminal-Bench 2.1 — GLM-5.2 81.0
Terminal-Bench 2.1 — Claude Opus 4.8 85.0
Open weights (MIT), MoE architecture, up to 1M-token context. Artificial Analysis Intelligence Index ≈ 51 — within striking distance of leading closed models.
What it signals
If disk streaming of MoE experts can be made fast enough, the rule that huge models demand huge memory could loosen considerably.
Conceptually striking
A signpost toward cheaper local deployment of frontier-class open weights.
Still a demo
Bottlenecked by speed; SSD wear and heat are open concerns. Not a GPU replacement — yet.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…