JetBrains ships Junie Local mode with bundled Qwen3.6-27B for M5 Macs with 64 GB RAM

The headline change is operational, not a new model release: Junie has long been able to point at Ollama or LM Studio, but the user had to pick a model, write a profile and accept whatever the local small model could or could not do. Junie Local collapses that setup into a single `/local` command that downloads the weights, starts the local server and switches the agent over. Plan mode, live prompting, guidelines, skills and `/commands` all behave the same as before, only the engine underneath changes. For an engineer already using Junie, the on-ramp is one command rather than a configuration exercise.

The hardware floor is the gating cost of adoption. JetBrains calls out M5 specifically because the M5 Neural Accelerator has 8-bit arithmetic instructions that M4 lacks, and using them in MLX yielded about 40% more prefill throughput, most of an agent's wall-clock time is prefill, reading files, not generation. The team is sending that change upstream as a PR to MLX-VLM. On top of that they reuse the KV cache across tasks and run speculative decoding, which together roughly double generation speed. None of this changes the headline constraint: an M5-class Mac with 64 GB of unified memory and roughly 20 GB of disk for the weights.

The model choice is deliberate and worth understanding. JetBrains picked Qwen3.6 over Qwen3.8 because 3.8 only works reliably with reasoning enabled, and with reasoning on, task throughput drops by roughly a factor of four. They ship Junie Local with reasoning disabled entirely, since their own tests found it added little quality at two to three times the token cost. That is a structural trade-off: any task where chain-of-thought actually moves the needle will look weaker locally than the same task against a reasoning-enabled cloud model, and that gap is most visible on complex architectural reasoning rather than day-to-day coding work.

On JetBrains' private coding eval the local 27B sits on par with Sonnet 4.5 at a 10,000-token reasoning limit, with GPT-5 at medium effort scoring slightly higher. Two caveats travel with those numbers: the local model is reasoning-off while the cloud comparisons had reasoning on, and the eval is JetBrains' own, useful as a reference point, not as an independent benchmark. For mechanical work, multi-file refactors, rename sweeps, test-coverage backfills, dependency upgrades, getting oriented in an inherited repo, the local agent is in the same conversation as the cloud ones it replaces.

What actually shifts is what engineers are willing to hand to the agent. Cost has been a dial in Junie for a while, plan on a strong model, implement on a cheap one, and Junie Local turns that dial to zero. That unlocks work that was never worth the spend: long refactors, repetitive mechanical edits, sweeping test-coverage gaps. It also runs fully offline once the weights are on disk, which is the practical reason some teams will adopt it rather than the price. Where it stops working is exactly where JetBrains says it stops working: complex architectural reasoning still favours cloud models, and the M5-plus-64 GB requirement puts it out of reach for the majority of readers today, with broader hardware and a lower memory floor explicitly named as the next thing to push on.