On 16 October JetBrains said Junie Local supports NVIDIA RTX Spark from day one, and that Windows support has moved from experimental to stable. In their own test on RTX Spark, generation reached up to 55 tokens per second. They call that an initial performance baseline for the platform, with further inference optimizations already in development. The 55 is not an independent retest, and it is not a published benchmark suite.

[1]
A closed laptop, charcoal lid and muted-blue base
A closed laptop, the lid painted charcoal and the base muted blue, with blank key shapes along the side., AI-generated illustration, not a news photograph

After install, the local mode reads code, edits files, and runs tests. There are no token charges and no quotas, and the code never leaves the machine. The close of the post also says it is free, with no registration, no subscription, no credits, and no card. RTX Spark is described as a Blackwell GPU with a Grace CPU, in slim laptops and compact desktops. JetBrains thanks NVIDIA for early hardware and for the joint work to launch.

[1]

The context window is sized to the memory available to the GPU, so more memory keeps more code, tool output, and conversation. For inference they say they use Blackwell tensor cores for FP8 computation, and speculative decoding that drafts several tokens and verifies them together. That is their description of their own stack. They say they will share benchmarks as the later improvements reach Junie Local, so those later numbers are not in this post.

[1]

When Junie Local launched in August, it needed an M5 Mac with 64 GB of memory. In September they shipped experimental Windows support for NVIDIA RTX cards with 24 GB of VRAM. That support is now stable. RTX Spark needs Windows 11 on Arm and at least 24 GB of memory available to the GPU, running natively. The other path is Windows 11 on x64 with an Ampere-or-newer RTX GPU of at least 24 GB, and the post names the 3090, the 4090, and a 32 GB 5090. The driver must be NVIDIA 581 or newer, with about 20 GB of free disk for the model.

[1]

The model choice is Qwen3.6-27B or JetBrains’ merged Qwen3.8-3.6-27B-blend. They say the blend keeps much of Qwen3.8’s coding ability and Qwen3.6’s shorter reasoning. More models are still to come. The installer recognizes the PC, checks the GPU and driver, and installs an inference engine, the model, and the integration. The install command itself is not in this draft.

[1]