JetBrains releases Junie Local, an offline coding agent

JetBrains released Junie Local on August 24, a coding agent inside its IDE that runs entirely on your own hardware once you type /local, per JetBrains's launch post. It pulls about 20GB of 4-bit weights, starts a local server, and after that nothing leaves the machine: no tokens, no quota, no bill. It's free, but the hardware bar is steep. You need an M5 Mac with 64GB of RAM, or this is a preview and not a tool you can use yet.

JetBrains says Qwen 3.6-27B scored close to Claude Sonnet 4.5 on its own internal eval, with GPT-5 at medium effort a step ahead. That's a vendor number on a vendor benchmark, so weigh it that way. More useful is the companion post on how they got a 27B model to run on a laptop at all, in JetBrains's engineering writeup. They tested Qwen 3.8 first and dropped it. With reasoning on, it produced 5x more tokens for a 4x slowdown, unusable on a laptop. So they moved to 3.6, turned reasoning off outright, went 4-bit over 8-bit, added 8-bit arithmetic during prefill, and layered multi-token prediction with n-gram speculative decoding on top. They also extended the rolling context so the agent's KV cache survives across tasks instead of re-reading the same files, and turned multi-agent mode off completely.

The timing outside JetBrains matters too. OpenAI reinstated its five-hour cap on ChatGPT Plus on August 25, exempting only the $100 and $200 tiers, according to a thread on r/OpenAI. A free Chrome extension called Diet Claude, built to track and dodge Claude's own rate limits, took the number two spot on Product Hunt. Junie Local doesn't negotiate a better limit. It removes the limit, for anyone with the hardware to run it.

I don't have an M5 with 64GB of RAM. Nobody's published what a full workday inside Junie Local feels like once the novelty wears off and multi-agent mode stays off.