JetBrains has released Junie Local, a version of its AI coding agent that runs entirely on your Mac with no cloud inference, no token metering, and no code ever sent to a server. The move marks the first time the company has shipped agentic AI that works completely on-device.
Installation is frictionless as all you need to do is open Junie and run the /local command, the model downloads, the local server starts, and Junie switches over. There is no JSON profile to write, no runtime to install first, and no endpoint to configure. Your existing setup, Plan mode, live prompting, guidelines, skills, and custom commands, all carry over unchanged; only the inference engine switched.
You Need an M5 Mac with 64GB of RAM
Junie Local requires an M5 Mac with a minimum of 64GB of RAM. The model itself, Qwen 3.6-27B at 4-bit quantization, demands approximately 20GB to download; JetBrains does not shy away from the constraint. An M5 Mac with 64GB “is a big ask” and puts the tool out of reach for most developers.
The M5 chip is not arbitrary, the M5 Neural Accelerator includes 8-bit arithmetic instructions that the M4 lacks, which JetBrains leveraged to achieve roughly 40% more prefill throughput, the phase where the model reads your source files to understand context. That performance jump was essential to making a 27-billion-parameter model respond fast enough to feel interactive on local hardware. The company is already prototyping versions for discrete GPUs, including DGX Spark and RTX 5090, with plans to explore 24GB cards.
Why JetBrains Chose Qwen 3.6 Over Newer Models
JetBrains deliberately selected Qwen 3.6-27B over the newer Qwen 3.8 because the newer model requires reasoning to be enabled in order to work reliably, and with reasoning on, tasks run approximately four times slower. Qwen 3.6 wins on speed, and on Mac hardware today, speed is the constraint that matters most.
The company also disabled reasoning entirely in the local model. Internal testing found that reasoning added very little quality improvement while consuming two to three times as many tokens. The result: Qwen 3.6-27B with reasoning off scores roughly on par with Claude Sonnet 4.5 (which uses a 10,000-token reasoning limit) and slightly below GPT-5 at medium effort. For everyday coding work, that gap is often invisible; on complex architectural reasoning, it is not.
When Inference Costs Nothing, Behavior Changes
Work that never justified the spend on a metered service suddenly becomes rational to hand off to an agent when iterations are free. Junie Local is built for:
- Multifile refactors and renames: Tasks not worth the token cost before.
- Test coverage gaps: Tedious work teams have ignored for quarters.
- Dependency upgrades and migrations: Repetitive mechanical tasks agents handle well.
- Repository orientation: Understanding an unfamiliar codebase without spending cloud credits.
Long, repetitive, mechanical work is exactly what an agent is designed for; cost-zero iterations remove the economic friction that once made such requests irrational.
Code Never Leaves Your Machine
For teams operating under client NDAs or in regulated industries, Junie Local removes the vendor-risk conversation entirely. There is no data policy to review, no third-party provider in the chain, and no code ever leaving your hardware. Once the model weights are on disk, Junie Local works exactly the same way on an airplane as it does at your desk, no network required.
How JetBrains Made Inference Fast Enough
The conventional approach to benchmarking AI models measures token generation speed. For a coding agent, that turned out to be the wrong metric; most execution time is spent in prefill, while the model reads files and understands the codebase. JetBrains optimized prefill instead, using KV-cache reuse across tasks and speculative decoding that roughly doubles generation speed. The M5 Neural Accelerator’s 8-bit arithmetic instructions were critical, yielding that 40% prefill throughput boost. JetBrains is upstreaming these optimizations as a pull request to MLX-VLM so other projects can benefit.
Mac First, But Not the Final Stop
JetBrains started narrow because tuning a single model across the entire stack beats supporting every model poorly; Mac was the starting point, not the destination. The company already has working prototypes for DGX Spark and RTX 5090, with 24GB cards in exploration. Prefill behavior differs substantially on discrete GPUs, so much of the optimization work will shift to that hardware. Junie Local is completely free, no registration, no subscription, no credits, no card required. For developers with M5 Macs and the RAM to match, the trade-off is straightforward: accept the hardware ceiling and gain unlimited, private, offline inference. For everyone else, JetBrains’ recent shift of Junie’s default to Gemini 3.7 Flash at 40% off base pricing offers a cloud-based middle ground.
