JetBrains Releases Junie Local, Bringing Its Coding Agent Fully On-Device to Macs

JetBrains has released Junie Local, an on-device version of its AI coding agent that runs entirely on supported Macs without cloud inference or token charges. The tool uses a JetBrains-tuned, 4-bit quantized Qwen3.6-27B model, which the company says scored on par with Claude Sonnet 4.5 on its private coding benchmark, with GPT-5 at medium effort scoring slightly higher. Junie Local requires a Mac with Apple's M5 chip and at least 64 GB of unified memory, with the least expensive compatible configuration priced at $4,499.

TL;DR — Key Takeaways - JetBrains has launched Junie Local, an on-device version of its AI coding agent that runs entirely on supported Macs without cloud inference or token charges. - Junie Local uses a JetBrains-tuned, 4-bit quantized Qwen3.6-27B model and can inspect code, make changes, run commands and iterate locally. - JetBrains says its internal testing put Qwen3.6-27B roughly on par with Claude Sonnet 4.5 on its private coding benchmark, with GPT-5 at medium effort scoring slightly higher. Mind you, you’ll need one heck of a Mac to run it. JetBrains has released Junie Local https://blog.jetbrains.com/junie/2026/08/junie-local-launch/ , a version of its AI coding agent that runs entirely on a user’s Mac. That means no cloud inference, no need to transmit source code to external model providers, and, pay attention, this is the important bit, no token charges. Junie https://www.jetbrains.com/junie/ , for those who haven’t used it, is JetBrains https://www.jetbrains.com/ ’ AI coding agent: software that can take a development task, inspect a project, make multi-file changes, use development tools, run tests, and iterate rather than only suggest code in a chat window. It’s available inside JetBrains IDEs and via Junie CLI in a terminal. This agent can work in JetBrains IDEs, from the terminal, and in headless CI/CD use cases. Previously, the tool relied on configurations involving external model providers or user-managed endpoints. The company’s CLI also supports custom model profiles for compatible endpoints, including local providers such as Ollama https://ollama.com/ . But those setups require developers to choose, install, and configure a model and runtime themselves. JetBrains is pitching Junie Local as a more turnkey alternative. The company bundles its chosen model, a local inference server, and an agent-oriented profile designed to support the repeated reasoning-and-tool-use cycles that distinguish coding agents from ordinary chat-based code completion. Junie Local packages Junie’s agent workflow with a locally managed, JetBrains-tuned Qwen 3.6-27B https://huggingface.co/Qwen/Qwen3.6-27B model. With Junie Local, you can make code changes, execute commands, and complete tasks iteratively without sending prompts, repository content, or tool output off the device. That distinction matters because agentic coding workloads are demanding. The system must retain context across multiple files, select and invoke tools reliably, interpret command output, edit code, and recover from failed attempts. JetBrains said it selected and quantized Qwen 3.6-27B for this use case, rather than asking users to assemble their own local model stack. You may ask yourself why JetBrains didn’t pick the newer, better Qwen 3.8 https://qwen.ai/blog?id=qwen3.8 . Dmitry Savelev, Junie’s head of marketing, explained, “Qwen3.8 needs reasoning enabled to work reliably, and with it on, tasks run roughly four times slower. On today’s Macs, 3.6 wins.” How good is it? According to Savelev, “We evaluate every model on JetBrains’ own private test set before it goes near Junie, and Junie Local was no exception. Qwen3.6-27B scored on par with Sonnet 4.5 10,000-token reasoning limit . GPT-5 at medium effort scored slightly higher.” If those comparisons are right, SWE-bench’s verified test reports https://www.swebench.com/verified.html show Sonnet 4.5 at 77.2%, while GPT-5 scored 74.9%. That would put Junie Local in the frontier-capable programming range. That’s pretty darn impressive. But you can only run Junie Local on Macs with Apple’s M5 chip with at least 64 GB of unified memory. Those are not cheap. The least expensive Mac configuration that met those requirements I could find was a 16-inch MacBook Pro with an M5 Pro, 64 GB of unified memory, and a 2 TB SSD https://www.apple.com/shop/buy-mac/macbook-pro/16-inch-space-black-standard-display-apple-m5-pro-chip-18-core-cpu-20-core-gpu-64gb-memory-2tb-storage for $4,499. Ouch If you want more horsepower, say a MacBook with M5 and 128 GB of RAM, you’re looking at $6,999, and you won’t be getting your machine until early October. But hey, you won’t need to pay for tokens, and that’s not a small matter. According to new Gartner research, AI coding agent spending will surpass the average software developer’s salary https://devops.com/ai-coding-costs-could-exceed-developer-salaries-gartner-warns/ by 2028. These Macs are starting to look affordable. Junie Local itself requires no subscription, AI credits, account registration, or payment for local operation. Of course, once you get yourself into any toolchain, the bills will eventually appear for additional software and services. There’s no such thing as a free lunch. Another audience that would probably like Junie Local is those of you programming in regulated industries, security-sensitive organizations, and companies with strict code-governance rules. For them, Junie Local’s big selling point is its data control. In the local configuration, code and prompts stay on the user’s machine rather than being sent to cloud model providers. The trade-off, of course, is the hardware costs. JetBrains said the broader effort aims to enable fully local operation across a wider range of hardware configurations. Still, it has not detailed a timeline for Linux, Windows, or lower-memory Mac support. I don’t see it running on lower-powered hardware. But we’ll see what JetBrains can come up with. As Savelev wrote, “We know that an M5 Mac with 64 GB of RAM is a big ask. We are not going to pretend otherwise, and we know it puts Junie Local out of reach for many people reading the post. That is simply what it costs to run a 27B model well today, and it is the number we are working hardest to bring down. The aim is a lower memory floor, wider hardware, and more of the stack optimized. If the lofty requirements are the reason you cannot try Junie Local, rest assured that we are working to bring them down.” Well, I wish them luck.