For the episode 14 of my series in which I test local models that fit on a 16GB MacBook Pro I chose Alpie Seed, a post-train of Qwen 3.5 9B. It comes from a company called 169pi, headquartered in Mumbai. They focus on edge computing and local models. The total size of the model was about 5.6 GB.

It’s a medium level model, both in terms of speed and performance. The “what is a dolphin?” question was answered in about 35 seconds, at roughly 6tok/s, which is not very fast. The answer rather short, other models had 3x-4x more text on this one.

The coding task (same word-count utility) took under three minutes and gave a working result plus basic usage notes.

It’s usable for day-to-day office work, summarization, and small coding jobs.

Among the models tested so far it sits in the middle. Fine for tiny projects. Not something I’d use to refactor a real codebase or start a larger app from scratch.

Some Models Are Better Than Others

Like I said, this is the 14th video in this series, which means I got 14 testing sessions with 14 different models. A few patterns are starting to emerge, and I can start seeing a video in which I try to put everything together.

The leading platoon is Gemma and Qwen, with the medium layer comprised of LFM and / or post-trains of Qwen, or newcomers like K2 Horizon (trained from scratch) or even Apple Intelligence.

But beyond individual performance, the thing that is getting more and more important is the fact that most of these models are actually usable, with a little bit of effort and fine tuning, on consumer hardware. Keep in mind that all the data stays on your computer, and you don’t pay for API tokens.

As usual, if you like this kind of content, like, share and subscribe to my YouTube channel.