Overview
- Junie Local is available inside Junie via a single /local command that downloads about 20 GB of model weights and starts a local server that switches the agent to run entirely on the developer’s machine.
- Once the weights are stored, all inference happens offline so prompts, source and diffs never leave the device, offering a privacy-first, unmetered workflow for air-gapped or NDA-bound work.
- JetBrains ships a preconfigured Qwen3.6-27B model quantized to 4-bit with reasoning disabled to cut token use and latency; internal tests on JetBrains’ private suite show Qwen3.6 near Sonnet 4.5 while GPT-5 scored slightly higher.
- The runtime uses an mlx-vlm inference engine built on Apple’s MLX to exploit M5 Neural Accelerator instructions, which JetBrains says improved prefill throughput by about 40 percent because reading repo context is the main bottleneck.
- The release imposes a high hardware floor—macOS 26 on an Apple M5-class machine with roughly 64 GB unified memory—so JetBrains calls this a first, narrow step and says it will work to lower memory needs and add support for other hardware while enabling free, long, mechanical developer tasks like large refactors and migrations.