JetBrains launches Junie Local, a free offline coding agent for high-end Macs

JetBrains has introduced Junie Local, a free, entirely offline variant of its AI coding agent that runs directly on a developer's machine. The move addresses growing demand among engineers to keep source code private and avoid per-request API costs by processing everything locally.

One-click simplicity for local inference#

Unlike existing approaches where developers must separately install runtimes like Ollama or LM Studio and manually select a compatible model, Junie Local handles the entire setup within the IDE. A developer simply selects "Junie Local" from the model menu, triggers an automatic download (about 20 GB), and the agent immediately switches to running the bundled inference engine. No configuration, no separate services, no model selection—everything is pre-optimized.

The release comes after JetBrains' broader Junie agent became available in standalone CLI form earlier this year. Developers have been able to point Junie at local model runtimes for months, but the new Local offering eliminates that friction entirely.

Why Qwen3.6, and the M5 Mac problem#

JetBrains deliberately selected Qwen3.6-27B—a 27-billion-parameter model from April—over the newer Qwen3.8-27B that arrived in August. The company's head of marketing notes that Qwen3.8 requires its reasoning mode for reliable agent performance, but reasoning multiplies task completion time by roughly four-fold. On current Macs, "3.6 wins," according to the announcement.

The model runs at 4-bit quantization using an inference engine built on mlx-vlm, which leverages Apple's MLX machine-learning framework. This approach mirrors Ollama's own pivot to MLX in March, maximizing the unified-memory advantages of Apple Silicon chips.

However, the hardware floor is steep: the setup demands macOS 26, at least 64 GB of unified memory, and an Apple M5 chip or newer—effectively limiting Junie Local to MacBook Pro M5 Pro or M5 Max tiers. JetBrains openly acknowledges the barrier but frames it as the current cost of running a 27-billion-parameter model efficiently, promising work to reduce those requirements in future iterations.

Where the real work happens#

A surprising finding shaped JetBrains' optimization focus: most agent latency comes not from generating tokens but from the prefill stage—the time spent loading source files, prompts, and context before the model begins producing output. Rather than chasing faster generation speeds like traditional benchmarks suggest, JetBrains optimized the prefill path, which is where meaningful gains matter most for coding workflows.