The newer model lost. JetBrains picked version 3.6 over the fresher 3.8, and said why.
There isn't one model behind local Junie. There's a list of trade-offs, made one at a time, until the agent fit on a laptop.
Fact
The facts, by JetBrains' own numbers: memory from previous tasks is kept, so already-opened files in the same project don't get re-read. The system prompt and the initial context stay remembered across sessions. Reasoning is turned off, giving 2 times faster execution for a small quality loss. Optional model calls have been stripped out. The model is quantized to 4 bits instead of 8, giving roughly 2 times faster generation at the same prefill speed. Prefill itself runs at 8 bits on M5 chips and speeds up by about 40 percent. On top of that sits multi-token prediction plus matching on repeated text, up to 2 times faster generation. Source: the JetBrains blog.
Why newer comes out worse here
Qwen 3.8 is the stronger model on paper. Only it needs reasoning turned on, and at medium effort produces around five times more tokens than 3.6. The result is roughly four times slower work. On a machine with limited memory, that isn't a trade-off. It's a brake.
On a local machine, newer is sometimes worse.
The list is instructive beyond JetBrains, too. None of these moves is new on its own. The gain comes from stacking them. If someone promises you a laptop agent just because they swapped models, you know what's missing.
Official primary sources
JetBrains Blog: Qwen for JunieThe visual is generated code-art, no third-party images.