← Retour au blog
tech 1 September 2026

Running Qwen3.8-Flash-Next on a 48GB Mac: A Technological Feat

Learn how to run a massive 104GB model on a Mac with just 48GB of RAM, thanks to ingenious use of SSD storage. A feat that could redefine the accessibility of advanced AI.

Article inspired by the original source
Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s ↗ github.com

Introduction

In the world of artificial intelligence, the scale and power of models continue to grow. However, this rapid advancement poses a significant challenge: the hardware resources required to run these models are becoming increasingly demanding. Today, we will explore how an innovative developer managed to run the Qwen3.8-Flash-Next model, weighing 104GB, on a Mac equipped with just 48GB of RAM. How? By leveraging the power of SSD streaming.

The Challenge of Large Models

Natural language models like Qwen3.8-Flash-Next are often designed to run on robust computing infrastructures, often beyond the reach of developers and small businesses. With a size of 104GB, this model typically requires a machine with at least as much RAM, making it inaccessible to many.

The Solution: Streaming from SSD

Carlos Lfu, the developer behind 'slotstream', proposed an ingenious solution: use SSD storage to stream the model's necessary data rather than loading it all into RAM. This approach utilizes MLX and Swift technology, making the model compatible with the Ollama API. In simple terms, it means that only the parts of the model that are needed at a given moment are loaded into memory, while the rest is stored and accessible on the SSD.

Performance and Results

Using this method, Carlos managed to achieve an impressive throughput of 12 tokens per second on a 48GB Mac. This is not only a victory in terms of technical feasibility but also opens the door to democratizing access to advanced AI models.

Impact on the Industry

This innovation has the potential to transform how developers and businesses interact with AI models. By reducing hardware barriers, it allows more players to participate in AI's evolution. It may also encourage further innovations in optimizing and effectively utilizing resources.

Conclusion

Carlos Lfu's 'slotstream' project is a striking example of how ingenuity and innovation can overcome apparent hardware limitations. It also shows that the future of AI does not solely rely on increasing hardware capabilities but also on improving methods to use them effectively.

Let's discuss your project in 15 minutes.

AI models SSD streaming Qwen3.8-Flash-Next Mac hardware innovation
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call