Junie now runs entirely offline. Can you spare a 64 GB M5 Mac?
The New Stack Paul Sawers
JetBrains made Junie run fully offline on your Mac. It’s free, but it wants an M5 with 64 GB of memory, so this is not for bargain laptops.
Based on reporting by The New Stack, Paul Sawers — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Most AI coding tools still lean on cloud models, but JetBrains is pushing Junie in a different direction. The company has launched Junie Local, a free version of its coding agent that runs entirely on the developer’s own machine, with no separate model host in the middle and no internet connection required once it’s set up.
That setup is meant to be painless. Inside Junie, the /local command downloads the model and inference engine, starts a local server, and flips the agent over automatically. There’s no Ollama or LM Studio installation to juggle, no endpoint to wire up, and no model profile to build by hand. JetBrains is also choosing the model for the user here, rather than leaving them to piece together the stack themselves.
The model is Qwen3.6-27B, an open-weight 27-billion-parameter release from April. JetBrains looked at the newer Qwen3.8-27B too, but decided the older one behaved better inside Junie on current Macs. Savelev says Qwen3.8 needed reasoning mode switched on to work reliably with the agent, and that made tasks take about four times longer. So for now, JetBrains is betting on the slightly older model because it is the one that actually works well enough in this setup.
The catch is the machine it wants. JetBrains says Junie Local needs about 20 GB of downloads, macOS 26, at least 64 GB of unified memory, and an Apple M5 chip or newer. In practical terms, that points at M5 Pro or M5 Max MacBook Pro machines. JetBrains is not pretending that’s a small ask. Savelev says the company knows the hardware bar is high and is working to lower it.
The reason for that bar is not just raw generation speed. JetBrains says the real bottleneck for a coding agent is prefill, the part where the model reads files, prompts, and context before it starts producing output. That makes Junie Local feel aimed at the jobs people hate paying attention to anyway: long refactors, renames, dependency upgrades, framework migrations, and test coverage work. And because it’s free and unmetered, developers don’t have to keep glancing at token usage every few minutes.
Privacy is the other obvious draw. Once the download is done, prompts, source, and diffs stay on the machine, with no outside model provider sitting between the developer and the code. For proprietary projects, client work, or anything under a strict NDA, that’s the real selling point, not the novelty of another local model wrapper.
My take — AI-written commentary, not fact-checked reporting
This is the right move, and the hardware cliff is the honest part. Local AI tools keep getting sold as freedom, but most of them still quietly demand a pile of setup work and a tolerance for compromise. JetBrains is at least admitting the trade: if you want the code to stay put, you may also need a very expensive Mac to keep it there.
Read more about this at: The New Stack
