Meet Colibrì: Running a 1.5-TB LLM on Just 25GB of RAM

Italian engineer Vincenzo—known online as JustVugg—has quietly dropped what might be the most intriguing local AI proof-of-concept of the year. Dubbed Colibrì, it's a system that manages to load and run GLM-5.2, a 744-billion-parameter Mixture-of-Experts model weighing in at a hefty 1.5 TB, on hardware that would make most AI engineers laugh: a modest CPU, 25 GB of RAM, and a 1 GB/s virtual NVMe drive.

Yes, you read that right. A 1.5-trillion-parameter-class model, squeezed into a setup that costs less than a mid-range smartphone. The catch? It's glacially slow—clocking in at 0.05 to 0.1 tokens per second. A single question takes hours. But that's not really the point.

How Colibrì Cheats the Laws of AI Physics

The secret sauce behind Colibrì is brutally simple in concept and fiendishly difficult to execute well: it loads the model in slices to RAM, swapping layers in and out as needed. Instead of keeping the entire 1.5 TB model resident in memory—which would require a small fortune in hardware—Colibrì pages through the model on demand, trading speed for memory efficiency.

This isn't entirely new territory. Projects like llama.cpp have long offered CPU-based inference and memory swapping. What sets Colibrì apart is the scale: it targets a genuine frontier-level MoE model, not a distilled 7B or 13B variant. GLM-5.2, developed by the Tsinghua-affiliated Zhipu AI team, sits somewhere in viewing distance of the best offerings from Anthropic, OpenAI, and Google on capability benchmarks.

Vincenzo's own early testing reportedly produced "impressive results" in terms of answer quality—the kind of depth you'd expect from a cloud-based flagship model, running on what is essentially a souped-up home computer.

What This Means for Local AI

The implications ripple well beyond one Italian engineer's living room. Colibrì suggests a future where running cutting-edge LLMs locally doesn't require an Nvidia NVL72 rack or a data center connection. For enthusiasts and organizations concerned about data privacy, subscription costs, or API dependency, that's a significant signal.

Currently, the local AI scene is split between small quantized models (4-bit 7B/13B) that run comfortably on consumer hardware and cloud-only behemoths that require enterprise budgets. Colibrì's approach points to a third path: massive models, limited memory, and patience as a design parameter. If the technique can be optimized—and that's a big if—the ceiling lifts dramatically for what's possible on a desktop machine.

The Hardware Reality Check

Let's be direct about the limitations:

• Speed: At 0.05–0.1 tokens/s, Colibrì is unusable for real-time conversation. The threshold for natural interaction sits around 20–30 tokens/s.
• Scaling: Higher-end hardware improves throughput but still falls short of practical use today.
• Bandwidth bottleneck: The 1 GB/s NVMe drive is the primary constraint—faster storage would improve swap speed significantly.

But here's the thing: none of these limitations are fundamental laws. Faster consumer storage (PCIe 5.0 NVMe drives already hit 10+ GB/s), smarter paging algorithms, and MoE-specific optimizations could each deliver order-of-magnitude improvements. If even two of those three factors improve, we're looking at a viable local inference path for models that currently require cloud APIs.

A First Step, Not a Finish Line

Vincenzo himself frames Colibrì as exactly what it is: a proof-of-concept. The project is open-source, available on GitHub, and already attracting attention from researchers and hobbyists alike. The fundamental idea—that you can run a 1.5 TB model on 25 GB of RAM by treating memory as a sliding window rather than a fixed pool—is the kind of creative engineering the AI hardware scene desperately needs right now.

With AI data centers consuming ever more power and memory prices climbing, approaches that decouple model quality from hardware requirements become increasingly valuable. Colibrì won't replace your ChatGPT subscription tomorrow. But it opens a door that many assumed was welded shut.

And for a proof-of-concept, that's exactly the right thing to do.