Overview
- JetBrains released Junie Local on Monday, August 24, 2026, and the agent installs inside Junie with a single /local command that downloads the runtime and switches the tool to an offline mode.
- Junie Local ships a quantized Qwen3.6‑27B model in 4‑bit form that requires roughly 20 GB of download space and an Apple M5 Mac running macOS 26 with about 64 GB of unified memory to operate.
- To speed real‑world coding work JetBrains tuned an mlx‑vlm/MLX inference engine for prefill throughput on the M5 Neural Accelerator, reporting about 40% higher prefill rates versus earlier chips.
- The shipped build has model reasoning disabled to cut token cost and speed execution, which JetBrains says narrows capability on complex architectural reasoning but is sufficient for many everyday and repetitive coding tasks.
- JetBrains calls this a first step and is working to lower the 64 GB memory floor and add support for discrete GPUs such as DGX Spark and RTX 5090, a shift that could broaden who can run private, unmetered agents on local hardware.