Junie Local is now stable on Windows and RTX Spark: requirements, install command and which model to pick

JetBrains announced on October 16 that Junie Local, the version of its Junie coding agent that runs a model entirely on your own machine, now supports NVIDIA’s RTX Spark PCs from day one. Windows support also moves from experimental to stable. If you have a 24 GB NVIDIA card or one of the new RTX Spark laptops, the practical questions are whether your machine qualifies, how to install it, and which of the two bundled models to use.

Key facts:

• Price. Junie Local is free: no registration, subscription, credits or card. JetBrains says the model download is the only network activity, and after that prompts, source code and diffs stay on the machine. (Source: Junie Local)
• Speed on RTX Spark. JetBrains measured generation speeds of up to 55 tokens per second and calls that an initial baseline. Junie Local uses FP8 on Blackwell Tensor Cores and speculative decoding. (Source: JetBrains)
• Context. The installer picks the largest context window your GPU memory can hold, so more memory means more code, tool output and history in context.

Does your machine qualify?

JetBrains lists two supported Windows setups:

PC	What you need
NVIDIA RTX Spark	Windows 11 on Arm, at least 24 GB of memory available to the GPU
Windows PC with NVIDIA RTX	Windows 11 x64 and an RTX GPU with at least 24 GB of VRAM

You also need NVIDIA driver 581 or newer and about 20 GB of free disk space for the model. On desktops, JetBrains says any Ampere-or-newer 24 GB card works, naming the RTX 3090, 4090 and 32 GB 5090. On macOS the requirement is still an Apple Silicon M5 Mac with 64 GB of RAM, as at the August launch.

Install

On Windows, run this in PowerShell (the command JetBrains publishes):

powershell -NoProfile -ExecutionPolicy Bypass -Command "[Environment]::SetEnvironmentVariable('JUNIE_LOCAL_MODEL', '1', 'Process'); iex (irm 'https://junie.jetbrains.com/install.ps1')"

Then start Junie, run /model, and pick a model. Junie downloads what it needs, starts the local engine and switches your session over. On a Mac, the product page gives curl -fsSL https://junie.jetbrains.com/install.sh | bash -s -- --local-model, and /local inside Junie does the same switch. Your guidelines, skills and /commands carry over, and /model switches you back to a cloud model.

Since the September preview, JetBrains says it rebuilt the installer and local engine. The installer now checks your GPU and driver before installing the inference engine and model, and the engine starts, stops and updates more reliably in long sessions.

Which model to pick

You choose between two models:

• Qwen3.6-27B, which JetBrains recommends for interactive iteration on simpler tasks. It ships with reasoning disabled. On JetBrains’ internal benchmark it scored 29.5 (±2.5), against 29 for Sonnet 4.5 and 33 for GPT-5. Those cloud models had reasoning on. If you’d rather run a Qwen3.6 coding model yourself, our Qwen3.6-35B-A3B local coding guide covers llama.cpp quants and VRAM.
• Qwen3.8-3.6-27B-blend, an equal-weight merge of Qwen3.6-27B and Qwen3.8-27B with no extra post-training, for longer and harder tasks. In JetBrains’ September write-up, it solved 37 of 100 internal coding tasks, versus 34 for Qwen3.6 with reasoning off and 39 for Qwen3.8. It generated 71% fewer output tokens than Qwen3.8.

One caveat from JetBrains: the blend can still overthink when it is stuck. If Junie keeps revisiting the same approach without new evidence, interrupt it and restart with a narrower goal.

If you want a different model

Junie Local only ships those two tuned models. JetBrains says Junie can also connect directly to Ollama, LM Studio and LiteLLM through interactive setup, with no JSON profile needed, so you can bring your own local model. JetBrains says more models and lower memory requirements are in progress, but hasn’t given dates.