JetBrains has shipped Junie Local, a version of its AI coding agent that runs entirely on your machine. No tokens, no quota, no subscription, and no code ever leaving your hardware. It is free, and installing it takes one command: type /local inside Junie, and the model downloads, a local server starts, and the agent switches over.
The move matters because it attacks the three complaints that have kept local AI coding agents niche until now: setup friction, model selection, and quality. Junie already let you point it at Ollama or LM Studio and bring your own local model, but that meant picking the model, tuning settings, and accepting that small models only handle simple tasks. Junie Local removes all of that. It ships a model JetBrains chose and tuned against its own agent loop, so the decision making is done for you.
Here is what shipped, what is under the hood, and the honest catch in the fine print.
What You Get
Inside Junie, you run /local. There is no JSON profile to write, no runtime to install first, and no endpoint to configure. About 20 GB downloads, and everything after that happens on your hardware, so your prompts, source code, and diffs stay put.
The model is Qwen3.6-27B at 4-bit quantization. The hardware requirement is specific: an Apple M5 Mac with 64 GB of RAM. That bar immediately rules out most machines sitting on most desks today, which we will come back to.
Your existing Junie setup carries over unchanged. Plan mode, live prompting, guidelines, skills, and your /commands all behave the same way. As JetBrains puts it, the engine changed, the agent did not.
And the price is the part worth repeating: completely free, with no registration, no subscription, no credits, and no card required.
The Interesting Engineering: Prefill, Not Generation
The deep dive published alongside the launch contains the most instructive part of this story. Everyone benchmarks generation speed, the tokens-per-second number that fills local LLM benchmark charts. For a coding agent, that turns out to be the wrong number to chase.
A coding agent spends most of its time on prefill, reading files to work out what is going on in your codebase. On an M5 Mac out of the box, prefill ran at roughly 650 tokens per second, versus something like 3,700 tokens per second on an RTX 5090. When the model requests file contents, the majority of wait time is prefill, not generation.
JetBrains dug into why prefill was slow and found something odd: prefill speed was identical across 4-bit, 8-bit, and 16-bit quantization. The reason is that prefill is compute-bound, and most matrix operations during prefill were running in full 16-bit precision regardless of how the weights were stored. The M5's Neural Accelerator has 8-bit arithmetic instructions that the M4 lacks. By patching MLX-VLM to switch matrix operations during prefill to 8-bit on the self-attention layers, JetBrains got around 40% more prefill throughput. The company plans to submit the patch upstream as a PR, and the same idea can be applied in vLLM by editing a model config file.
That instruction set detail is also why Junie Local starts at M5 and not M4. The M4's 16-bit arithmetic delivers 20 to 30% slower prefill, and it cannot use the 8-bit trick at all.
There is more in the stack of tricks. Junie now adds every new request directly to the agent's rolling context, so KV-caches from previous tasks get reused instead of forcing the model to re-read files it has already seen. Speculative decoding combining multi-token prediction with n-gram matching roughly doubles generation speed.
Why Qwen 3.6 and Not the Newer 3.8
The model choice looks backwards at first. Qwen3.8 is newer, so why ship 3.6?
Because Qwen3.8 needs reasoning mode enabled to work reliably. Without it, quality degrades badly enough that the model can get stuck repeating the same tool call in a loop. With reasoning on, roughly 5x more tokens get produced, which nets out to about 4x slower tasks. On today's Macs, the older model wins.
Junie Local ships with reasoning disabled entirely. JetBrains found that reasoning added very little quality while costing two to three times as many tokens, which translates to roughly a 2x speedup in task execution. The 4-bit quantization choice follows similar logic: it benchmarks only slightly worse than 8-bit, and since generation is memory-bound, 4-bit is about twice as fast.
How Good Is It?
JetBrains evaluates every model on its own private test set before it goes near Junie. On that set, Qwen3.6-27B scored on par with Sonnet 4.5 running with a 10,000-token reasoning limit, and GPT-5 at medium effort scored slightly higher.
Read those numbers with the right frame. The local model ran with reasoning disabled, against cloud models that had reasoning switched on. And the test set is JetBrains' own, not a public benchmark, so you cannot independently reproduce the comparison. For everyday work, JetBrains' claim is that you would likely not notice the gap, and on complex architecture-heavy tasks the cloud models still pull ahead.
A useful piece of outside context: JetBrains' own Developer Ecosystem Survey 2026 found that 90% of professional developers now use AI coding agents at work weekly. An on-device option with zero per-token cost is a natural next step for that audience, especially for anyone working on code they are contractually or legally not allowed to send to a cloud API.
The Catch: The Hardware Bar
Everything above sounds like a giveaway, and it is, with one condition. You need an M5 Mac with 64 GB of RAM. That is a high-specced, expensive machine, not the base configuration anyone buys. If you are on an M-series Mac with 16 or 24 GB of unified memory, or on Windows or Linux, Junie Local is not for you yet.
The 20 GB download also means you are committing real disk space, and the free price tag arguably serves a second purpose: it makes the hardware the only barrier, which pushes demand toward exactly the machines Apple sells at the highest margins. That is speculation, but the alignment is hard to miss.
What Comes Next
Mac was the starting point, not the plan. JetBrains says it already has working prototypes for NVIDIA's DGX Spark and the RTX 5090, and is looking at 24 GB graphics cards. Prefill behaves very differently on discrete GPUs, where it is far less of a bottleneck, so much of the optimization work will shift as the hardware does.
The Takeaway
- Junie Local runs JetBrains' AI coding agent entirely on your Mac, free, with no account or credits required
- Setup is one command,
/local, with about 20 GB of downloads and no configuration - It runs Qwen3.6-27B at 4-bit, tuned against Junie's agent loop, with reasoning disabled
- The real performance story is prefill optimization, not generation speed, including an 8-bit prefill patch worth about 40% more throughput on M5
- JetBrains rates it on par with Sonnet 4.5 on its private test set, with GPT-5 medium slightly ahead
- The barrier is hardware: an M5 Mac with 64 GB of RAM, with Windows, Linux, and discrete GPU support still in prototype
If you already own a maxed-out M5 Mac, this is an easy thing to try this week. For everyone else, it is a signal worth watching: the first serious attempt by a major IDE vendor to make on-device coding agents a default option rather than a tinkerer's project.
Based on JetBrains' announcements: Junie Can Now Run Entirely on Your Mac and How We Optimized the Qwen 3.6 Model for Our Junie Agent.
Top comments (0)