JetBrains launches Junie Local, a version of its Junie AI coding agent that runs entirely on-device on Apple M5 Macs with 64GB RAM, using a tuned Qwen3.6-27B model at 4-bit quantization. Installation is a single /local command with no configuration, tokens, quota, or cloud dependency, keeping code fully on the machine. The model benchmarks near Sonnet 4.5 and slightly below GPT-5 at medium effort. JetBrains chose Qwen3.6 over 3.8 for speed, optimized prefill throughput using M5's 8-bit Neural Accelerator instructions (about 40% faster), and used speculative decoding to roughly double generation speed. Support for DGX Spark and RTX 5090 is in prototype.
Table of contents
Our first step into on-device coding agentsOne command, no configurationThings we learned about making it fastHow good is it?This is the first stepGetting startedQuestions this post answers
What are the hardware requirements to run Junie Local from JetBrains?
Junie Local requires an Apple M5 Mac with 64 GB of RAM, since the local model needs about 20 GB of storage for its weights. The agent runs a 4-bit quantized version of Qwen3.6-27B tuned specifically for JetBrains' Junie agent loop, entirely on-device with no cloud connection or token usage required. daily.dev helps developers weighing local versus cloud AI coding setups track releases like this one.
Why did JetBrains choose Qwen3.6 instead of Qwen3.8 for Junie Local?
Qwen3.6 was chosen over the newer Qwen3.8 because Qwen3.8 requires reasoning enabled to work reliably, and enabling reasoning makes it run roughly four times slower. On current Mac hardware, Qwen3.6 with reasoning disabled offers a better speed-to-quality tradeoff for a local coding agent. Developers choosing between model versions for local agents can follow these tradeoffs on daily.dev.
How does Junie Local's performance compare to cloud models like Sonnet 4.5 and GPT-5?
Qwen3.6-27B running locally in Junie Local scored on par with Sonnet 4.5 at a 10,000-token reasoning limit, while GPT-5 at medium effort scored slightly higher. Notably, the local model achieves this with reasoning disabled entirely, since testing found reasoning added little quality while costing two to three times as many tokens. daily.dev keeps developers evaluating local versus cloud coding agents up to date on benchmarks like these.