Junie Local, JetBrains’ AI-powered coding assistant, has been available from day one on the NVIDIA RTX Spark platform, while its Windows support for RTX cards with at least 24 gigabytes of video memory has moved to the stable stage. This enables the agent to run locally to read code, edit files, and run tests without the code leaving the device.
RTX Spark relies on a GPU based on the Blackwell architecture alongside a Grace processor, with the aim of running larger local models on thin laptops and small desktop computers. According to JetBrains’ initial tests on the platform, Junie Local reached a generation speed of 55 tokens per second, with additional inference improvements still in progress.
What changes in practice?
Junie Local automatically adapts the context-window size to the memory available on the GPU. The more memory available, the more code, tool outputs, and conversation history the agent can retain during a task. It also takes advantage of Blackwell’s Tensor Cores to perform FP8 calculations efficiently, and uses speculative decoding to generate several tokens in advance and then verify them together.
According to JetBrains, the local experience includes neither token charges nor usage quotas, and does not require a subscription, account registration, or payment card. However, performance and the ability to retain a long context remain tied to the available hardware memory.
Windows support and requirements
When Junie Local launched in August, it required a Mac computer with an M5 chip and 64 gigabytes of memory. In September, JetBrains launched experimental support for NVIDIA RTX cards with 24 gigabytes of VRAM, before announcing its stability now.
The software currently runs on NVIDIA RTX Spark devices with Windows on Arm, and on Windows x64 computers equipped with RTX cards based on the Ampere architecture or newer, with at least 24 gigabytes of video memory, such as the RTX 3090, RTX 4090, and RTX 5090 with 32 gigabytes. Running it also requires NVIDIA driver version 581 or newer and approximately 20 gigabytes of free storage space.
Models and how to use them
The installer automatically rebuilds the device configuration, checks the GPU and driver, then installs the optimized inference engine, model, and Junie integration. Users can choose between Qwen3.6-27B and Qwen3.8-3.6-27B-blend, a blended model that JetBrains says combines a large portion of Qwen3.8’s coding capabilities with shorter reasoning than Qwen3.6.
To begin installation on Windows, JetBrains provides a PowerShell command. After launching Junie, the /model command can be used to select the model. The software then downloads the required components, runs the local engine, and switches the session to it.
Limitations and next steps
Local support does not mean that all cards or models will provide the same experience; context size and generation speed depend on memory and hardware. The figure of 55 tokens per second also represents an initial baseline rather than a final performance benchmark. JetBrains says it is working on adding more models, improving coding performance, reducing the time and memory required to complete tasks, exploring more efficient model formats, and improving speculative decoding on RTX Spark.