September 18, 2026
Junie Local saves tokens on code: blend model uses 71% fewer than Qwen3.8
On September 18, JetBrains added a blend model to Junie Local that generated 71% fewer output tokens than Qwen3.8 across 100 internal coding tasks. Qwen3.8-3.6-27B-blend solved 37 tasks, versus 39 for Qwen3.8 and 34 for Qwen3.6 without reasoning. The update also brought experimental Windows support in a nightly build.
JetBrains
@jetbrains
Junie has a new home on X: @junie_ai 👋 Follow for updates on new features and announcements. Starting with a smarter Junie Local, a new blend model, and experimental Windows support. The team shared the details below 👇
We mixed Qwen 3.8 and 3.6. Literally. A 50/50 weight merge. One 27B model. 71% fewer output tokens than 3.8 in our internal coding eval. Junie Local has an update. Windows devs, you're invited too 🤝
· 24.1K views
Junie Local ran on macOS 26+ with Apple Silicon M5 or newer, 64 GB of RAM, and about 21 GB of free space. This setup now adds an experimental Windows build for NVIDIA RTX on Ampere architecture or newer with at least 24 GB of VRAM.
Blended weights. In `Qwen3.8-3.6-27B-blend`, JetBrains combined the weights of Qwen3.8 and Qwen3.6 at a 50/50 ratio without additional post-training. Across 100 tasks in its internal coding benchmark, the blend solved 37 tasks, versus 34 for Qwen3.6 without reasoning and 39 for Qwen3.8.
Across 30 tasks solved by both the blend and Qwen3.8, the blend generated 279 thousand output tokens versus 935 thousand for Qwen3.8. It used fewer tokens in 29 of the 30 tasks.
Quality. Across four LiveCodeBench runs, the blend averaged 85,47% correct answers, versus 83,29% for Qwen3.8 and 67,87% for Qwen3.6. On a MacBook Pro with M5, speculative decoding with two tokens accelerated decoding by 60%, while with four tokens the gain fell to 36%.
Getting started. Junie works inside the IDE and in the terminal. In an active session, run `/local`, select a model, and Junie will download the inference engine and weights, configure a local server, and switch the agent to it.
For the new blend model, update Junie, run `/local`, and select `Qwen3.8-3.6-27B-blend`. On Windows, the nightly build launches with `junie --channel=nightly`.
After downloading the weights, the agent works without a network connection. Junie Local is free: no registration, subscription, AI credits, or bank card required.
JetBrains states Zero Data Retention by default and does not train models on code. Junie remains a proprietary product rather than open source.
Windows support remains in the nightly build for now.
Source
