JetBrains’ Junie Local update aims to make local coding agents both smarter and faster, capable of understanding a codebase and finishing tasks on a user’s machine without long waits.
The update centers on making a more capable model practical to run locally, without the previous trade-off between two distinct versions of the same model.
FROM SPEED VS. REASONING TO A UNIFIED BLEND
Previously, running Qwen3.6 with reasoning disabled yielded usable speed on laptops, while Qwen3.8 with reasoning enabled could complete more tasks but took longer. The new release blends the two in equal proportions to form a single 27B model, removing the need to choose between speed and reasoning capability. TechStaged has also covered OpenAI uses coding agents to reshape AI research, with early data on usage and acceleration.
WHAT THE BENCHMARKS SAY ABOUT THE BLEND
In internal coding benchmarks, the blend completed more tasks than Qwen3.6 with reasoning disabled and approached the performance of Qwen3.8, while delivering a substantial token savings. For example, in a 100-task evaluation, the blend completed 37 tasks, compared with 34 for Qwen3.6 and 39 for Qwen3.8, and it generated 71% fewer output tokens than Qwen3.8.
EFFICIENCY AND TOKEN SAVINGS AT TASK SCALE
Across a subset of 30 tasks tested against Qwen3.8, the blend used roughly 70% fewer tokens (about 279K vs 935K) on those tasks.
Longer coding sessions showed token efficiency was nuanced: while the blend generally used fewer tokens than Qwen3.6, it could spend more time on reasoning, which affected total token usage in some benchmarks.
LIVECODEBENCH RESULTS AND HARDWARE NOTES
The blend averaged 85.47% correct answers across four LiveCodeBench runs, slightly higher than Qwen3.8 (83.29%), with Qwen3.6 trailing at 67.87%. The 27B blend also required far fewer tokens per complete pass than Qwen3.6 (about 6.14 million vs 24.1 million).
MODEL DESIGN AND RUNTIME OPTIMIZATIONS
A multi-token prediction (MTP) head accelerates decoding by proposing multiple tokens per round. In on-device tests, using two tokens per round significantly sped up decoding on an M5 MacBook Pro, while four tokens per round offered diminishing returns due to added GPU work. The blend uses a four-bit (Q4) MTP head by default, which kept memory usage lower without sacrificing accuracy.
WHY THE BLEND CAN REQUIRE FEWER TOKENS DESPITE LONGER THINKING
The team found that the extra tokens in some cases came from the reasoning path, with longer traces contributing to token counts. They observed that selecting shorter, correct traces could cut token usage by about 24.5% without harming performance.
SEED HANDLING AND REPRODUCIBILITY IN TESTING
Early internal tests used a fixed random seed per request, which created numeric instability and biased the evaluation. The team corrected this by advancing the seed with each agent step, allowing different exploration paths while preserving reproducibility in testing.
AVAILABILITY AND HOW TO TRY THE BLEND
Apple M5 users can try the new model via Junie by running /local and installing the blend. For Windows, a nightly Junie build includes experimental RTX support for NVIDIA cards based on Ampere or newer with at least 24 GB of VRAM. The blend is published on Hugging Face as Qwen3.8-3.6-27B-blend.
WHAT’S NEXT FOR JUNIE LOCAL
This blend is one result of an ongoing set of model and runtime experiments. JetBrains notes that more results from this line of research will appear in future Junie Local releases.
RELATED COVERAGE
- OpenAI uses coding agents to reshape AI research, with early data on usage and acceleration
- More Capable, Affordable AI Expands What Work Can Be Accomplished
- Gemini 3.8 Flash arrives in Junie with 75% launch discount, tailored for long-horizon coding tasks
- From Leaderboards to Model Profiles: JetBrains Ties 523 Tasks Across LLMs While Highlighting Divergent Coding Trajectories
- Software articles
SOURCES
- The JetBrains Blog: Making Local AI Smarter and Faster Published · Primary source
