Local Qwen and some magic Skills

Qwen 3.8 27B

I have been using the Qwen 3.8 27B MLX based model on my local machine and it's amazing how good local models are getting. It can spend a large number of tokens reasoning, but the output from this model is really good so far.

The model I'm currently using is Jundot/Qwen3.8-27B-oQ4e-mtp. I'm running this with oMLX on a Mac Studio.

Matt Pocock skills

I've also been spending some time learning to use skills to help with software work locally. I was messing with Superpowers, but found it too broad. It's better to focus on a smaller change at a time. It just didn't feel right for me.

I recently found some videos discussing skills by Matt Pocock: Skills For Real Engineers. The skills I have used in this set have been great. They are focused on specific problems and do a good job at keeping you focused on the current task.

• grill me and grill-with-docs are both great.
• handoff lets you postpone work for later without losing any context you already have.
• teach builds lessons so you can learn almost anything.
• improve-codebase-architecture helps you with big architectural changes.

Every one of the tools I have tried has done a great job. You can find videos related to these skills and other topics on his YouTube.