Italian Engineer Successfully Runs a 744-Billion-Parameter AI on a Regular PC in 2026 – Affordable Local AI Solution

Running massive AI models is usually associated with expensive servers, high-end GPUs, and operating costs that are far from budget-friendly. However, in 2026, an Italian engineer demonstrated a different approach: a 744-billion-parameter AI model can apparently run on a regular PC through a local solution called Colibrì. Although its performance is still far from ideal, this achievement opens a new path for AI computing that is more affordable, private, and not entirely dependent on the cloud.

What Is Colibrì and How Does It Work?

Colibrì is experimental software designed to enable extremely large language models to run on home computers. Its main focus is not speed, but proving that inference with massive models is still possible without data center infrastructure.

Colibrì Software for Loading the 1.5 TB GLM-5.2 Model on a Home Computer

One of the most striking things about Colibrì is its ability to load the GLM-5.2 model, which is around 1.5 TB in size. This is clearly too large to fit entirely into the RAM of a regular PC. Because of that, Colibrì uses a staged loading approach and leverages NVMe storage as high-speed virtual memory.

With this method, a home computer does not need hundreds of gigabytes of RAM or a GPU with massive VRAM. The system only retrieves the parts of the model needed as the inference process runs. Technically, this approach does sacrifice speed, but it gives ordinary users a chance to run models that were previously only realistic in data centers.

Mixture-of-Experts (MoE) Architecture as the Key to Efficiency

Another key behind this experiment is the use of the Mixture-of-Experts (MoE) architecture. Unlike regular dense models that activate all parameters for every token, MoE activates only some of the relevant “experts” at each step.

This means that even though the model has a total of 744 billion parameters, not all of them are working at the same time when generating an answer. This is what makes ultra-large models more feasible to run on much simpler hardware. Its efficiency does not mean it is fully lightweight, but it is enough to reduce the computational barrier compared to dense models of equivalent size.

PC Specifications & Performance Challenges Faced

This achievement is interesting, but it is important to understand it realistically: “can run” does not always mean “comfortable to use.” Colibrì is still currently at the proof-of-concept stage.

Minimum Configuration: Standard CPU, 25 GB RAM, and 1 GB/s Virtual NVMe

This experiment is said to run on a relatively affordable configuration: a standard CPU, around 25 GB of RAM, and virtual NVMe storage with a speed of about 1 GB/s. This is far lower than the requirements of conventional AI servers, which usually demand data-center-class GPUs and large amounts of memory.

For many users, those specifications are still fairly reasonable for a modern desktop PC or home workstation. This is where Colibrì becomes appealing: it shifts the idea that massive AI models can only exist on expensive infrastructure.

Extremely Slow Speed (0.05–0.1 Tokens/Second) & No GPU Support Yet

The biggest challenge lies in performance. The reported speed is still extremely slow, at around 0.05–0.1 tokens per second. In practice, this means a single response could take a very long time, especially if the requested answer is fairly long.

In addition, Colibrì is also said not to support GPUs yet. As a result, the entire process depends heavily on the CPU and the mechanism for fetching data from storage. Until major optimizations are made, using it for real-time chatbots is still impractical.

The Prospects of Local AI: Benefits, Privacy, and Cost

Although slow, the idea behind Colibrì has major implications for the future of local AI. Many users do not always need ultra-fast responses, especially if their priorities are privacy, data control, and cost savings.

A Cost-Effective Alternative for Users Concerned About Privacy & Subscription Fees

Local AI offers an important advantage: data does not need to be sent to third-party servers. For users handling sensitive documents, internal research, or personal needs, this approach feels safer and more reassuring.

In addition, local models also have the potential to reduce dependence on monthly subscription fees. If technologies like Colibrì continue to mature, users could have their own AI system at home without having to keep paying for premium cloud access.

Proof-of-Concept Status & Future Optimization Steps

For now, Colibrì is more appropriately viewed as a proof-of-concept than a ready-to-use solution. Its greatest value lies in proving that technical barriers can be overcome with creative approaches, even if the user experience is not yet ideal.

The next optimization steps will likely focus on GPU support, more efficient memory management, faster weight streaming techniques, and adjustments to drastically reduce latency. If these areas continue to develop, it is entirely possible that ultra-large local AI will become more practical in the next few years.

FAQ

Can a 744-billion-parameter AI model run on a regular laptop?

In theory, yes, but it depends heavily on the laptop’s specifications and the software implementation. In the context of Colibrì, “can run” refers more to technical proof than to a comfortable everyday user experience.

How long does it take to get a single answer from Colibrì?

Because its speed is only around 0.05–0.1 tokens per second, a single answer can take a very long time. The longer the requested response, the greater the waiting time.

What is the difference between Mixture-of-Experts architecture and regular AI models?

Regular models generally activate all parameters when processing input. Meanwhile, Mixture-of-Experts activates only some of the relevant “experts,” making it more efficient for extremely large models.

When can Colibrì be used practically for real-time chatbots?

Not anytime soon, based on its current performance. Colibrì will only become more realistic for real-time chatbots after major optimizations, especially in inference speed and GPU support.

Source: https://telset.id/news/ai/insinyur-italia-jalankan-model-ai-744-miliar-parameter-di-pc-biasa

This article was written by artificial intelligence (AI) using the deepseek-v4-pro model via SumoPod AI.

This article was translated by Artificial Intelligence (AI) using gpt-5.4 via SumoPod AI.