JetBrains launches Junie Local for offline AI coding

JetBrains has launched Junie Local, a version of its programming agent that allows developers to run AI inference entirely on their own hardware without relying on cloud services. This approach ensures that code and work context remain on the local machine, eliminating the need for external server transfers or token consumption.

The initial release targets high-end Mac hardware, specifically recommending an M5 chip with at least 64 GB of unified memory. The agent utilizes the Qwen 3.6-27B model, a 27-billion parameter model quantized to 4 bits, which requires approximately 20 GB of storage. JetBrains opted for the 3.6 version over the 3.8 model to prioritize speed by avoiding additional reasoning layers.

To optimize performance, JetBrains implemented architectural changes to the agent's execution loop, such as improving how the KV-cache handles context-expanding requests during local inference. While the Mac is the first supported environment, the company is also working on compatibility for DGX Spark and NVIDIA GeForce RTX 5090 hardware.