JetBrains announced Junie Local, a version of the Junie coding agent that runs entirely on a Mac instead of relying on cloud services. The feature does not use credits or paid quotas, and the company says that prompts, code, and diffs do not leave the device.
Junie previously supported connecting to local model runtimes such as Ollama and LM Studio, but this required users to choose the model, configure its settings, and manually prepare the runtime environment. Junie Local, by contrast, comes with a model selected by JetBrains and configured to work with the agent loop. It can be activated from within Junie using the /local command. The model is then downloaded, the local server starts, and the agent switches to local operation without requiring users to write a JSON file, install a separate runtime environment, or specify an endpoint.
High Requirements and Local Operation
Junie Local uses the Qwen3.6-27B model at 4-bit precision and requires a download of about 20 GB. JetBrains also requires a Mac equipped with an M5 processor and 64 GB of RAM. Once the download is complete, processing takes place on the local hardware, while the user's existing settings remain active, including planning mode, direct prompts, guidelines, skills, and /commands.
Why Did JetBrains Choose the M5 Processor?
JetBrains says that code-generation speed is not the most important metric for a coding agent; much of the time may instead be spent reading files and understanding repository context, a stage known as prefill. The team therefore focused on improving the performance of this stage, taking advantage of 8-bit computation instructions in the Neural Accelerator within the M5, which the company says are not available in the M4. According to JetBrains' measurements, this resulted in an increase of approximately 40% in the prefill processing rate, with plans to submit the fix to the MLX-VLM project as a merge request.
The company chose Qwen3.6 instead of the newer Qwen3.8 because the newer release requires reasoning to be enabled to work reliably, making tasks about four times slower according to its tests. JetBrains also used KV-cache reuse between tasks and a speculative-generation configuration that it says approximately doubles generation speed.
Performance and Practical Limitations
In a test suite developed by JetBrains, Qwen3.6-27B achieved performance comparable to Sonnet 4.5 within an inference limit of 10,000 tokens, while GPT-5 with a medium-effort setting scored slightly higher. However, the comparison is not entirely equivalent: the local model runs with reasoning fully disabled because JetBrains' tests found that it added little quality while consuming two to three times as many tokens, whereas the feature was enabled in the cloud models used for comparison. As a result, the difference may not be apparent in everyday work, but it becomes clear in complex architectural tasks.
What Changes Practically for Developers?
Eliminating cost accounting may make tasks that developers previously postponed or avoided more feasible for the agent to perform, such as refactoring and renaming across multiple files, closing test-coverage gaps, upgrading dependencies, carrying out framework migrations, and understanding a legacy repository. Local operation also provides a direct benefit to teams working with code subject to client confidentiality agreements, because the data does not enter a cloud-provider loop.
Nevertheless, Junie Local remains constrained by high hardware requirements and a large download size, and its performance on complex architectural reasoning is lower than that of cloud models according to the figures presented by JetBrains. The available source text ends before completing its explanation of the privacy implications, so the final details concerning usage policies and availability require human review.