A coding agent that never leaves the laptop, if it's an M5

JetBrains has released Junie Local, a free version of its coding agent that runs entirely on the developer's own machine, provided that machine carries at least 64 GB of unified memory and an Apple M5 chip. The launch was reported by The New Stack, based on a blog post published Monday by JetBrains' head of marketing, Dmitry Savelev.

At a glance

• Typing /local inside Junie downloads the model and inference engine, starts a local server and switches the agent over, removing the need to install Ollama or LM Studio or configure an endpoint.
• JetBrains picked Qwen3.6-27B, an open-weight model released in April, and runs it at 4-bit through an inference engine based on mlx-vlm, with the setup pulling around 20 GB of downloads.
• Because inference is free and unmetered, JetBrains positions the agent at long mechanical jobs such as multi-file refactors, test-coverage gaps, dependency upgrades and framework migrations, where token spend usually deters delegation.

Local coding agents have until now been an assembly job: choose a model, choose a quantization, wire up a runtime and hope the combination holds up under tool use. Shipping the whole stack pre-tuned reads as a bet that packaging, rather than raw model quality, is what has kept local agents marginal. The price of that bet is a hardware floor that likely excludes a large share of the developers the product would otherwise attract.

A /local command replaces the Ollama and LM Studio setup work

JetBrains launched Junie in January 2025 as an agent embedded in IDEs such as IntelliJ IDEA, PyCharm and WebStorm, able to plan tasks, modify code, run tests and inspections and work with a project's context, and later extended it into a standalone CLI. Developers could already point Junie at runtimes such as Ollama or LM Studio and load a model of their choosing.

Junie Local removes that assembly. JetBrains selected the model, quantized it and tuned the inference engine and agent harness around that specific combination, with Junie Local appearing in the model selector next to the cloud-hosted options. Cline, Continue and Aider can also be pointed at local runtimes, and GitHub added local-model support to Copilot CLI in April, including an offline mode for air-gapped setups.

Reasoning mode made Qwen3.8 roughly four times slower inside Junie

Junie Local uses Qwen3.6-27B, a 27-billion-parameter open-weight model released in April, rather than Qwen3.8-27B, which arrived earlier in August with improvements. Savelev writes that Qwen3.8 needed its reasoning mode enabled to work reliably with the agent, and that with reasoning switched on, tasks took roughly four times longer.

On today's Macs, 3.6 wins.

The model runs at 4-bit through an inference engine based on mlx-vlm, which in turn uses MLX, Apple's machine learning framework for Apple Silicon. Ollama took a similar route in March, when it moved its Apple Silicon engine onto MLX to take advantage of the unified memory architecture of the chips.

JetBrains says the optimization target was prefill rather than generation speed. Savelev writes that most of an agent's time goes on reading files to work out what is going on, which makes tokens per second the wrong number to chase, and that optimizing prefill produced the real gains.

The floor is macOS 26, an M5 chip and 64 GB of unified memory

Junie Local requires macOS 26, at least 64 GB of unified memory and an Apple M5 chip or newer, and the setup involves about 20 GB of downloads. On MacBook Pro, the memory figure puts buyers into M5 Pro or M5 Max configurations.

We know that an M5 Mac with 64 GB of RAM is a big ask. We are not going to pretend otherwise.

Savelev describes that requirement as what it currently costs to run a 27B model well, and as the number JetBrains is working hardest to reduce. He writes that everything after the download happens on the user's hardware, so prompts, source and diffs stay put.

Local execution addresses a familiar constraint for developers working on proprietary code, under client NDAs, or in settings where sending source to a third party is off the table. Savelev expects little practical gap against stronger cloud models on everyday work, while conceding that complex architectural reasoning remains better suited to them.

Which Macs come next

JetBrains states an intention to lower the memory requirement, support a wider range of hardware and keep optimizing the underlying stack, with Savelev telling developers blocked by the current specifications that the numbers should come down. No timeline, no target memory figure and no wider hardware list has been named, and Junie Local for now runs only on Apple Silicon under macOS 26.