JetBrains Releases Junie Local, a Fully Offline AI Coding Agent

JetBrains has launched Junie Local, a variant of its Junie coding agent that does all its work on the user's device, with no internet connection and no code, prompts, or output sent to external servers. It answers a growing demand from companies and developers working with sensitive code for AI tools that don't require trusting a vendor's cloud.

How local mode works

Starting Junie Local is as simple as typing the /local command in the plugin's interface. The tool downloads the model itself, spins up a local inference server built on the MLX engine, and switches the agent into offline mode, with no JSON config files to write, no separate runtime to install, and no endpoint to point to manually. Existing project settings, Plan mode, live prompting, guidelines, skills, and custom commands all keep working unchanged.

Source code, prompts, and the agent's output never leave the machine. For companies under strict data security policies, or working in environments with limited network access, that's a practical difference from cloud-based coding agents that send repository fragments to external APIs.

Model choice and performance trade-offs

JetBrains' team chose Qwen3.6-27B over the newer Qwen3.8, because the latter requires an extra reasoning step that, in testing, slowed down coding tasks by roughly a factor of four. Rather than optimizing for raw token generation, the engineers focused on speeding up the prefill phase, meaning the loading and analysis of project files, using KV cache reuse and speculative decoding.

On the Apple M5, context-loading throughput is, according to JetBrains' internal measurements, about 40 percent higher than on the previous M4 generation. In comparative benchmarks, Qwen3.6-27B with a 10,000-token reasoning limit scored close to Claude Sonnet 4.5, while GPT-5 on medium compute effort came out slightly ahead.

No tokens, no quota, no code ever leaving your machine - JetBrains, Junie Local announcement

The hardware barrier

For now, Junie Local runs only on Macs with an Apple M5 chip and at least 64 GB of RAM, which in practice limits the tool to a narrow slice of the newest and most expensive Apple machines. The company confirmed it's working on versions for other hardware, testing prototypes on the NVIDIA DGX Spark, RTX 5090 cards, and GPUs with 24 GB of memory, which would eventually open the tool up to Windows and Linux users with powerful graphics cards.

JetBrains recommends Junie Local mainly for repetitive tasks that demand less creativity, such as code refactoring, writing tests, updating dependencies, and migrating between frameworks. For more complex, multi-step architectural problems, the company still points users to the cloud version of Junie, which has access to larger models.

What it means for developers

The launch fits into a broader trend of moving coding agents onto local devices, driven both by concerns over code privacy and a desire to avoid the rising subscription costs of cloud AI tools. For teams working on code covered by non-disclosure agreements or sector-specific regulations, the ability to run a fully functional agent with zero network traffic could be the deciding factor in choosing a tool.

At the same time, the hardware requirement of a Mac with an M5 and 64 GB of RAM means that, in the coming months, only a relatively small group of users will be able to use Junie Local. Wider adoption depends on how quickly JetBrains delivers the promised versions for NVIDIA graphics cards and hardware outside the Apple ecosystem.