JetBrains Junie runs local: no tokens, no code uploaded
Junie Local runs a coding agent entirely on an M5 Mac with Qwen3.6-27B. The real product is a compliance answer, not a benchmark score.
JetBrains shipped Junie Local this month: the same coding agent that lives in your IDE, running entirely on your Mac. No cloud round trip, no token meter, no code leaving the machine. You type /local in Junie, a model downloads, a local server starts, and the agent switches over. The benchmark numbers are fine. The reason to care is that "no code leaving the machine" is a sentence you can put in a client contract.
What actually happened
Per JetBrains, Junie Local runs Qwen3.6-27B at 4-bit, a roughly 20 GB download, with reasoning disabled. There is no separate runtime to install and no JSON profile to write — the /local command handles it. It is free: no registration, no subscription, no credits, no card.
The hardware bill is the catch. JetBrains specifies an M5 Mac with 64 GB of RAM and says plainly that this "is a big ask," which they are working to bring down. On their own internal testing, the local model scored on par with Sonnet 4.5 capped at a 10,000-token reasoning limit, and slightly below GPT-5 at medium effort — with the local model's reasoning off and the cloud models' reasoning on. Those are vendor-reported numbers on an internal suite, so treat them as a floor for feasibility, not a ranking.
Why a local coding agent matters for your business
It converts a policy problem into a hardware purchase. If you do client work under an NDA, handle regulated data, or have a customer whose security review bans third-party model providers, the honest answer used to be "we don't use agents on that repo." A 64 GB Mac is cheaper than losing the engagement.
Zero marginal cost changes what you delegate. When a run costs nothing, you stop rationing the boring work: backfilling tests, renaming across a package, mechanical migrations. Metered agents quietly push teams to hoard tokens for the interesting tasks, which are exactly the tasks a mid-size local model is worst at.
Keep the cloud model for the hard problems. This is a routing decision, not a replacement. Cheap local passes for bulk edits, a frontier model for architecture and gnarly debugging. Build the workflow so swapping either side is a config change.
Watch the spec, not the launch. Mac-only and 64 GB today. The story is whether that floor drops — that is what decides if local agents are a niche or a default.
Key takeaways
- Junie Local runs Qwen3.6-27B at 4-bit fully on-device, enabled with the
/localcommand - Requires an M5 Mac with 64 GB RAM and a ~20 GB model download; reasoning is disabled
- Free — no registration, subscription, credits, or card
- JetBrains' internal scoring puts it near Sonnet 4.5 at a 10k reasoning cap; vendor-reported, so verify on your own repo
- The practical win is data residency plus zero marginal cost, not a benchmark win
Your model should be a swappable part. We build dev and automation workflows where local and hosted models route by task, so a pricing change or a security review never stalls the team. See how we build it.
Sources: JetBrains Blog.
- #local-llm
- #coding-agents
- #jetbrains
- #data-residency
- #dev-tools
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Zimbra CVE-2026-73570: 274 breached, patch shipped July 20
An unauthenticated RCE in Zimbra's SNMP handling is being exploited in the wild. 274 confirmed compromises, ~8,200 servers still unpatched five weeks after the fix.
Read itX kills Nitter: your API integration is a permission, not a right
X Corp sent cease-and-desist letters to Nitter with a 24-hour deadline. If your product reads someone else's platform, that access can end in a day.
Read it