# Junie Local gives Windows developers a managed local coding agent > JetBrains has made Junie Local stable on supported Windows RTX hardware. Its CLI connects to a local model server, while developers still control repository access, commands and review. By BIG CHANGE Editorial Published: 2026-10-17T09:34:13.549Z Updated: 2026-10-17T09:34:13.549Z Canonical: https://bigchange.ai/blog/junie-local-windows-coding-agent-setup ![Conceptual charcoal illustration of an unbranded desktop computer tower and a wired keyboard resting on a work surface.](https://bigchange.ai/api/media/file/junie-local-workstation-concept-v1.png) Conceptual illustration of a local workstation and wired keyboard; no Junie installation or coding result is depicted. AI-generated illustration by BIG CHANGE. JetBrains made [Junie Local stable on Windows](https://junie.jetbrains.com/blog/junie-local-rtx-spark/) on October 16, after an experimental RTX release in September. It also says the coding agent supports NVIDIA RTX Spark PCs from the platform's launch. For developers with the specified hardware, the new route joins Junie CLI to a model server on their own machine and handles model installation and engine startup. That combination answers a practical question left open by a bare local model download: how does a coding agent connect a repository task to local inference? The model generates through Junie Local. Junie CLI still reads project files, proposes or makes edits, and may run commands for the task. Developers remain responsible for the repository, permissions, diffs and test results. ## The big change - **What changed:** Junie Local's Windows path moved from experimental to stable. Its installer checks the hardware and driver, installs the model and CUDA inference engine, and configures Junie CLI to use a local endpoint. - **Why it matters:** A developer with supported RTX hardware can give Junie a repository task without paying JetBrains per token for this local model. The agent's file and command actions still need review. - **What to watch:** JetBrains reports up to 55 generated tokens per second in its initial RTX Spark testing. That measures generation speed in the company's setup, not completed coding tasks, correctness or performance on every supported PC. ## The route from task to model The [Junie Local manual](https://junie.jetbrains.com/docs/junie-local.html), dated October 15, describes a managed local inference engine with an OpenAI-compatible API. On Windows it packages a CUDA-accelerated `llama.cpp` engine and 4-bit GGUF models. The installer writes a Junie CLI model profile pointing to `http://localhost:19239/v1/chat/completions`. When the local model is selected, Junie starts the engine and stops it after the last Junie instance using it exits. JetBrains offers Qwen3.6-27B and its Qwen3.8-3.6-27B blend. The local installation uses quantized builds, rather than the full-weight models linked from the manual. The two versions can be installed side by side and selected through `/model`. JetBrains says its setup chooses a context window that fits the available GPU memory; the exact capacity therefore depends on the machine. The Windows engine uses multi-token prediction from the main GGUF model for speculative decoding. Those mechanics explain where inference runs; they do not establish the quality of a proposed change. The [Junie CLI quickstart](https://junie.jetbrains.com/docs/junie-cli.html) describes the other side of the workflow. Start the CLI at a project's root, then submit a task in its terminal interface. A prompt can refer to project files or folders with `@`. The October 16 release says Junie can read code, edit files and run tests. Its CLI documentation says potentially sensitive actions, including most terminal commands and MCP tools, may require the user's approval depending on settings. Inspect that approval choice as carefully as the selected model. ## Set up a first repository task These steps follow JetBrains' documentation. BIG CHANGE did not install Junie Local or run this task. 1. **Check the machine.** The release lists Windows 11 on x64 with an Ampere-or-newer NVIDIA RTX GPU and at least 24 GB of VRAM. For RTX Spark, it lists Windows 11 on Arm and at least 24 GB of memory available to the GPU. Both paths require NVIDIA driver 581 or newer. The current manual also lists the Microsoft Visual C++ runtime and roughly 21 GB of free disk space. Check those prerequisites before downloading a model. 2. **Install the local route.** Use the Windows installer linked from the [Junie Local instructions](https://junie.jetbrains.com/docs/junie-local.html). JetBrains documents a PowerShell command that sets `JUNIE_LOCAL_MODEL=1` for the installer process. Review any downloaded script against the official page before executing it. The manual says the installer verifies download checksums and fetches roughly 17 GB for the Windows engine and model. 3. **Choose the model.** Open Junie CLI in the project directory and run `/local` in its terminal interface. The setup screen checks requirements and lets you install Qwen3.6-27B or the blend. The manual says `/local` on Windows is available in Junie CLI's terminal UI. If the standalone installer was used, select the installed model with `/model`. 4. **Give it a bounded task.** In a disposable repository or a branch you can review, ask Junie to inspect a specific file and make one small change with a relevant test. The expected output is a proposed or applied diff and any test output. Examine both before accepting the work. JetBrains says the first response may wait for model loading and project-context processing. A useful success check is that Junie CLI shows the local model selected, produces a reviewable diff for the requested file, and reports the command or test result that you can verify yourself. If installation stops, the manual says downloads resume and a rerun fetches missing artifacts. If the first prompt is slow, check whether the engine is still loading before interpreting it as a failed task. The documentation describes engine status and health commands through `serverctl` for further diagnosis. ## Local inference has a specific boundary JetBrains calls Junie Local free, with no registration, subscription, credits or token quota for this route. A developer still needs the specified GPU, disk space and electricity; external tools or services used by a task may have separate costs. The company's statement that code stays on the machine concerns the local-model path it describes. The CLI can read files, run commands and use configured tools, so a project owner should check tool permissions and network-connected integrations before handling sensitive repositories. BIG CHANGE has not independently audited Junie's network or telemetry behavior. The October 16 speed figure is also a vendor result. A measured coding comparison would need the model choice, hardware, repository, task, tests and observed diff. None of those results was generated by BIG CHANGE. This article maps the documented route so a developer can decide whether a controlled trial fits their machine and project. ## Sources & further reading - [Junie Local Comes to the NVIDIA RTX Spark](https://junie.jetbrains.com/blog/junie-local-rtx-spark/) (October 16, 2026). Windows stability, RTX Spark support, requirements and JetBrains' initial generation-speed claim. - [Junie Local manual](https://junie.jetbrains.com/docs/junie-local.html) (dated October 15, 2026; checked October 17). Engine, quantized models, installation, local endpoint and recovery details. - [Junie CLI quickstart](https://junie.jetbrains.com/docs/junie-cli.html) (checked October 17, 2026). Repository task flow and command approval controls. ## Sources - [Junie Local Comes to the NVIDIA RTX Spark](https://junie.jetbrains.com/blog/junie-local-rtx-spark/) - [Junie Local](https://junie.jetbrains.com/docs/junie-local.html) - [Junie CLI quickstart](https://junie.jetbrains.com/docs/junie-cli.html)