For the 9th episode of my series in which I test local AI models that fit on a 16GB MacBook Pro, I chose Ling 3.0 Tiny. It’s a 7.9 billion parameter mixture-of-experts model with 1.7 billion parameters active at any time. It comes from Inclusion AI (also referred to as Ant Group or Bai Ling), a lab based in Hangzhou and affiliated with Alibaba. I mention this because Alibaba is better known for the Qwen models, which I’ve covered in earlier tests in this series.

On a 16 GB MacBook Pro the model takes roughly 4–6 GB on disk. That leaves enough space for other applications, if you feel like, but I’d keep it tidy. It has a native 128K context window, but to conserve memory I kept the working context way lower — 32K.

The Standard Tests

I ran the same two checks I use for the rest of the series, through MLX Serve.

First, a plain text question (“what is a dolphin”). The model answered quickly and cleanly at 45 tokens per second. That is the highest generation speed I’ve measured so far, which puts Ling 3.0 Tiny on pole position.

Second, a small coding task: write a utility that counts words in files inside a folder. It requested tools twice — once to write the script, once to run it and verify — and produced a decent, working result without obvious bugs. Generation stayed just as fast during the whole process.

Overall it is a strong model for its size. Text generation is snappy and the coding output is usable. The only practical limit on 16 GB is context length. It works very well for short, clearly defined tasks. For longer or more open-ended work I would put a larger orchestrator model in front of it and only hand Ling well-specified pieces. Incidentally, this is the pattern I suggest across most of the small models I’ve tested on this hardware so far.

If you like this kind of content, like, share and subscribe to my YouTube channel!