The developers of Draw Things have launched Local Code, an open-source AI assistant that lets Mac users download models and put them to work on their own computers.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Local Code Launches as Free, Private AI Assistant Tool for Mac
- Thread starter MacRumors
- Start date
- Sort by post date
The developers of Draw Things have launched Local Code, an open-source AI assistant that lets Mac users download models and put them to work on their own computers.
Haiku_Oezu
macrumors 65816
Stop it, you’re going to make Sam Altman cry by hoarding all that DRAM in your computers like selfish little ghouls
That’s memory that could go to an OpenAI data center so it can help deafen people living in the suburbs, have you no shame?
That’s memory that could go to an OpenAI data center so it can help deafen people living in the suburbs, have you no shame?
j569mxr
macrumors member
there are a ton of options to work locally. what makes this one better than the others? what makes it worthy of an article vs others?
CheesePuff
macrumors 68000
Well yeah because the only open source models that are targeting frontier models are from China, Qwen 3.8 Max is very impressive, and if it's run locally why are you concerned that they're Chinese?
Notably, this is from the developer of https://drawthings.ai/ which is a well regarded, Mac native, and privacy focused AI image app.
Hi! Liu Liu here. Will try my best as this is not simple to clarify:
1. It is not yet another UI harness, it is a inference engine + harness + execution environment, all packaged in one app.
2. Our inference engine is optimized for Mac (the Draw Things one), and it prefills really fast.
3. The more interesting bit technically is the "execution environment", this ~600MiB app packaged ffmpeg, libreoffice, Python, node.js, Lean, plenty of other tools as "virtual command-line" tools inside the app, you don't need to give it full-disk access or allow it to execute any other programs on your Mac to do work.
4. Trust boundary: other harnesses requires full-disk access, and asks you to trust them (they will protect you from "bad other apps", but not themselves). Local Code lets Apple enforces App Sandbox on us, so you trust Apple to make sure Local Code won't "act bad". It is still useful because if you want to operates on some folders, Local Code asks you permission and that permission is enforced by Apple, not by us.
5. It is fully open-source: https://github.com/drawthingsai/local-code-community You don't need to trust what I said, you can ask your agent to validate each and every claim made here.
There are so many choices out there like you said, it is really a struggle to clarify things. I use it daily and I think it is pretty sweet. But again, I have a M5 Max 128G so I can drive the most powerful open-source model out there right now (DeepSeek 4.1 Flash 1017), not every Mac users are so lucky so their mileage with this software can vary.
ThisIsMike
macrumors 6502
This is - IF it’s as advertised and if it works well (I recognize those are two big ifs) - fantastic.
I’m very interested in reading others thoughts on this, but it seems like a huge win.
I’m extremely uncomfortable with aspects of the LLMs. The more alternatives the better, particularly if it keeps my data private and doesn’t require a subscription. This is also potentially absolutely something Apple should love, as it further supports their approach to running as much AI on device as possible, which in turn will require the hardware capable of doing so, and which will ultimately translate into more sales for the mothership.
I’m very interested in reading others thoughts on this, but it seems like a huge win.
I’m extremely uncomfortable with aspects of the LLMs. The more alternatives the better, particularly if it keeps my data private and doesn’t require a subscription. This is also potentially absolutely something Apple should love, as it further supports their approach to running as much AI on device as possible, which in turn will require the hardware capable of doing so, and which will ultimately translate into more sales for the mothership.
would you trust Meta over chinese and Qwen is that ? some workloads then this more than acceptable. See some AI declines in shares recently, well for some enterprises string a bunch of studios together and run this or Llama etc ..
Free AI in effect no subs, no openAI etc
Spanky Deluxe
macrumors 603
Because Meta's one isn't very good. Google's ones are decent but the ones from Qwen have been a step ahead for a while now. They're truly fantastic for the size. Not sure what the issue is with running a Chinese model to you but the alternatives are models from US companies that have cozied up to that nonce in the US White House so it's not like they can be 'trusted'.
I'm not sure what you're saying. A surprising amount of what data goes to China? Unless you specifically connect these models to the internet, they cannot communicate outside and even then, they don't 'phone home'. That's just not how they function. Now sure, the developer of this software could add in some kind of phone home ability but it's not the models doing that. And you can monitor that if you're paranoid with the likes of Little Snitch.
I think you're misunderstanding how these models work and run locally. They don't 'escape'. They don't leave your computer. It's just a bit of software that runs on your machine. These models are not sentient or alive.
Because a marketing statement that it is local does not necessarily mean that it is local.
You should monitor all of your outbound traffic. A surprising amount goes to China. Why?
serpico007
macrumors 6502
The more I stay away from AI, it creeps slowly into my life via employment and every day personal tasks. As much as I dislike the bad effects to society AI creates, I can't put my head in the sand like an ostrich and believe this will all disappear when the bubble bursts. Some of this stuff will still continue in some form or other. Ugh.
CheesePuff
macrumors 68000
There are plenty of ways to run local models, you don't need to use this...
Unsloth is better
Theres also Premsys on iOS
Theres also Premsys on iOS
Nope If you turn your network off, then I think you are good. I would however reboot before connecting the network back to make sure there are no queued network actions.
You monitor outbound connections using your firewall. In mine I added a rule to log all outbound connections.
I fully agree. AI has its plus’s and minus’s but overall one of the big pain points is privacy and data ownership. We have these powerful “AI chips” in most modern computers, we should be doing a lot more AI on device.
Prior methods of running local LLMs work well enough but local AI will never gain the footing it needs unless software makes it easy for average users to use local models. I’ll definitely be testing this out.
tim_apple
macrumors regular
You have no clue what you are talking about. A model, in this context, is just a big file with a bunch of floating point numbers in it. If you run it locally, it’s local.
I’m using them and they work amazingly well, but one concern about them being Chinese is even if they don’t have ill intent baked in (we don’t know for certain) and you run them locally, they can still behave a little suspiciously because of the data they were trained on: https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model
Liu Liu here. I don't know why it worth an article, but some comparisons:
Local Code v.s. LM Studio Bionic / Unsloth Desktop: Local Code is fully open-source: https://github.com/drawthingsai/local-code-community, much better tuned for the largest models available (DeepSeek Flash series); LM Studio Bionic supports more models, distributed outside of App Store and requires Full-Disk Access.
Local Code v.s. Ollama: Ollama supports more models, support serving models; Local Code doesn't need you to download another harness, or install any other tools (python environment, node.js, etc) to be productive, everything is included in the ~600MiB software.
Generally: 1. Local Code packages inference engine + harness + environment together so you don't need to download any other software to put it together; 2. At any time, you don't need to / cannot let Local Code have full-disk access or download other software from the internet to run it locally (these are sanctioned by Apple through App Sandbox mechanism enforced on Local Code); 3. Because it packages everything, out-of-the-box experience (if you are doing casual vibe-coding) is better: it can use node.js to check its javascript code, use web-browse-cli to snapshot and check the finished web page, and then use python to launch a HTTP server for you to try it out, all without reaching these software on your system (everything is packaged into the Local Code deliverable).
(Sorry for repeating myself in the post again, but trying to practice to see if there are ways to clarify this better).
Why do you care that it's Chinese (and, yes, Qwen models are also from China)? Models can't spy on you, you know! And if you suspect that any code it might produce has "back doors" or other such tin-foil-hat concerns, you can always examine said code before you do anything with it.
I currently use Ollama + Qwen Model (qwen2.5-coder:32b) + Aider on my Mac. It's not great with large HTML/Javascript creation/updates when compared to Gemini, but for smaller changes it's great. And it's free.
It looks like this tool is basically just a front-end integration of the above? Maybe not even Aider (since it mentions not being able to modify local files due to sandboxing)?
It looks like this tool is basically just a front-end integration of the above? Maybe not even Aider (since it mentions not being able to modify local files due to sandboxing)?
thats all it is, I run ollama on a much larger Linux laptop than my air and works but not as nice as this one
Populus
macrumors G3
This is the path Siri should take going forward: being a local assistant that just works.
At first, I had my doubts about spending the money on an M5 Mac with 32GB of RAM. However, each day I’m happier about that decision, as more and more local tools are arising.
And you may say: it will always be quicker and much better to work on cloud LLMs. And while that’s true, I don’t want Anthropic or OpenAI agents on my mac having access to my personal documents… And there’s the possibility that in the future, the cloud LLM subscriptions will get increasingly expensive. So I’m all in on the local LLM scene. Plus, you can work with models that aren’t censored.
The only caveat is that the current Local LLM software, such as Jan, Osaurus, Unsloth… is confusing and difficult to learn. Hopefully this local code is more straightforward to use
At first, I had my doubts about spending the money on an M5 Mac with 32GB of RAM. However, each day I’m happier about that decision, as more and more local tools are arising.
And you may say: it will always be quicker and much better to work on cloud LLMs. And while that’s true, I don’t want Anthropic or OpenAI agents on my mac having access to my personal documents… And there’s the possibility that in the future, the cloud LLM subscriptions will get increasingly expensive. So I’m all in on the local LLM scene. Plus, you can work with models that aren’t censored.
The only caveat is that the current Local LLM software, such as Jan, Osaurus, Unsloth… is confusing and difficult to learn. Hopefully this local code is more straightforward to use
It's interesting to see the breadth of difference in AI awareness in the comments, from "but it's a Chinese model!", to using tools that were frontier 3 years ago, to those using factory stacks now.
On my part, it feels like this is Yet Another UI Harness... Codex/Claude Code/Opencode/Open WebUI/ LM Studio Bionic/ omp / nanobot / and the other MLX centric ones. Its just like when we had a new opinionated framework release for every language just a few years back (but everyone's still using React).
On my part, it feels like this is Yet Another UI Harness... Codex/Claude Code/Opencode/Open WebUI/ LM Studio Bionic/ omp / nanobot / and the other MLX centric ones. Its just like when we had a new opinionated framework release for every language just a few years back (but everyone's still using React).
The way EVERYTHING in the supply chain is a security risk these days, you should be doing development work on a VM or container as a bare minimum. All it takes is one stray dependency on a "trusted" platform and your machine and the data on it is compromised.
everlast3434
macrumors 6502a
Awesome! Im going to install it, connect my bank accounts, investments, and health data and I'll report back!