JetBrains has announced Junie Local, a way to run local models on your machine without having to deal with the manual complexity that comes with installing and running local models.
Junie, if you don't know, is an LLM-agnostic AI coding agent similar to Anthropic's Claude Code and OpenAI's Codex. It's been "generally available" since June this year, and can do a lot of the stuff you'd expect from a typical AI agent like executing multi-file refactors and generating automated test coverage.
For a while now, it's been possible to run open weights models locally using runtime environments like Ollama or LM Studio.JetBrains argues that this process can be tedious because picking weights, tweaking settings, and setting up host endpoints manually takes effort, which is why Junie now has a/localcommand. Junie Local relies on Qwen3.6-27B at 4-bit (with reasoning switched off), which is downloaded when you first run the/localcommand.
JetBrainssaysthat you'd need an M5 Mac with at least 64GB of RAM to run Junie Local properly, because the M5 has 8-bit arithmetic instructions inside its Neural Accelerator that process prefilling (the initial phase where an agent reads project files to process context before writing code) roughly 40% faster than M4 chips. The company also said that it chose Qwen3.6 over the newer Qwen3.8 (released earlier this month) because the latter runs slower on "today's Macs" since it needs reasoning enabled.
For now, the only way to run Junie Local is on a Mac. JetBrains said that in its tests, it found Qwen3.6-27B in Junie Local scored "on par with Sonnet 4," which is fine for everyday, not-so-complex work.
Obviously, not everyone has a high-spec Apple machine lying around to run a 27B model well; even JetBrains acknowledges this. The company said that it already has "working prototypes" for NVIDIA's DGX Spark and RTX 5090 desktop graphics cards that may materialize in the future.