MacMini Bionic and MATLAB coding
I'm interested in chatting with anyone who is using Bionic or another agent to write MATLAB code. I have puny 6GB PC GPU RAM but large 64GB McMini RAM. What model are you using, what joys and woes have you had?
3 Comments
Time Descending
Our local Ai quickstart might help here: GitHub - matlab/matlab-local-agent-quickstart: Run a local AI coding agent against MATLAB®, with no code, data, or prompts sent off the machine. · GitHub
The models we've reccomended at these sizes follow some internal testing and we include some example benchmark problems that they should be able to solve matlab-local-agent-quickstart/demos at main · matlab/matlab-local-agent-quickstart · GitHub
This is just the first release (3 weeks ago) and we are interested in learning what you'd like to see here.
Happy to join this conversation! I have been using an M1 Ultra Mac Studio with 128GB RAM in local Agentic AI workflows with MATLAB, and it works quite well. I beleive that my approach should also work for a Mac Mini with 64GB RAM, but you should consider whether better sized models will work for your RAM size based on your application.
I am now using LM Studio Bionic (free plan), with entirely local models. I initially used the regular LM Studio and was doing a lot of configuration of MCP connections manually. During my experiments with local agentic workflows, LM Studio relesed the Bionic desktop app to make choosing, configuring and running models easier, as well as providing their own coding agent, and I have found that to work much better. I also have a subscription to Codex, and will switch back and forth between a Cloud-connected Codex session that makes local LLM calls to LM Studio and using the Codex CLI making local LLM calls to LM Studio.
The models I use today are Qwen3 Coder Next 4-bit (~45GB), Qwen3 VL 30B (~18GB), and Qwen3.8 27B (~16GB). All of these fit nicely into 128GB of RAM and still leave room for an interactive MATLAB desktop session connected by the MATLAB MCP server. All of these models support native tool calling. There are other Qwen Coder MLX models that you should consider if RAM is a limitation, such as Qwen3-Coder-30B-A3B-Instruct-MLX-4bit. It writes great MATLAB code as well. You can also forego the Qwen3.8 model if you are strictly doing coding projects. I still recommend the VL model if you have some image processing tasks - some of my workflows include extracting text/meaning from screenshots of UI and models, and the Qwen3 VL 30B does a decent job of that.
LM Studio does a great job of just choosing the right default settings for models in most cases. I also use it to contol some additional settings. All of my models have context set to 65,535, with the Contect Overflow set to use "Rolling Window." The Local Model API is configured to use Just-in-time model loading. That way, LM Studio only loads a VL supported model if the query requires it and unloads it when it doesn't.
Under integrations, I am using the native GitHub integration that comes with Bionic and of course a Manual MCP server configuration for MATLAB MCP Server.
One of the best "joys" of this implementation is a feature called "LM Link". This let's me use the Locally app on my iPhone to connect to my at-home Mac Studio and access my models and sessions, connected to MATLAB, from anywhere in the world. All while the execution stays entirely local to my at-home infrastructure.
If you're interested in more details about my configuration and benchmarks as compared to a Windows environment, check out this repo. It also has some apps you can use to switch between cloud-connected and local-only Codex sessions and perform some benchmarks. https://github.com/connectjeff/codex-launcher
Sign in to participate
Tags
