Junie Local's Hybrid AI Model Boosts Coding Efficiency
New blended model delivers smarter coding assistance with 71% fewer tokens
1Hybrid model breakthrough
Junie Local's new Qwen3.8-3.6-27B-blend combines two existing models in equal proportions, eliminating the previous tradeoff between speed and capability. In coding benchmarks, it completed 37 tasks compared to Qwen3.6's 34, approaching Qwen3.8's 39 completions. The key innovation is generating 71% fewer output tokens than Qwen3.8 while maintaining higher task completion rates. This represents a significant efficiency gain for developers running AI assistance locally.
2Token efficiency proven
Detailed analysis of 30 identical tasks showed the blend model consistently used fewer tokens than Qwen3.8, averaging 279K tokens versus 935K. The savings weren't just from failing faster - the blend model used fewer tokens on 29 of 30 comparable tasks. This efficiency comes from optimized reasoning patterns rather than simply truncating unsuccessful attempts. Developers benefit from faster completion times without sacrificing solution quality.
3Simple merge, complex results
The technical approach used equal weight merging of Qwen3.8-27B and Qwen3.6-27B models, leveraging their shared architecture. Surprisingly, this straightforward blending produced better-than-expected results without additional post-training. The team prioritized this approach after initial benchmarks showed consistent improvements across multiple task types. While more complex optimization methods exist, this solution delivered immediate practical benefits.
4Benchmark performance
Across four LiveCodeBench runs, the blend model averaged 85.47% correct answers versus Qwen3.8's 83.29%, while using approximately 6.14 million tokens per pass compared to Qwen3.6's 24.1 million. The visual benchmarks revealed different tradeoffs, with the blend model using more tokens than Qwen3.8 but substantially fewer than Qwen3.6 with reasoning enabled. This demonstrates context-dependent optimization opportunities.
5Runtime optimizations
Junie Local employs multi-token prediction (MTP) to accelerate generation, with a 60% speedup achieved by proposing two tokens per round on M5 MacBooks. The team optimized the balance between proposal quantity and GPU workload, settling on a 4-bit MTP head that maintains 63% acceptance rate while conserving memory. Attention calculations emerged as the primary bottleneck during verification, increasing from 8.4 to 40.2 ms per round in longer contexts.
6Reproducibility challenge
Early development uncovered an unexpected performance degradation tied to random seed reuse in evaluations. Fixed seeds caused numeric instability, particularly affecting Qwen3.8's performance. The solution involved advancing the seed with each agent step and reflection attempt, maintaining reproducibility while allowing divergent paths. This debugging process revealed model-specific vulnerabilities in the reasoning architecture.
7Platform expansion
The update brings Junie Local to Windows machines through experimental NVIDIA RTX support, requiring Ampere or newer architecture with ≥24GB VRAM. Apple M5 users can immediately switch to the new model via the local installation command. This expansion significantly broadens the potential user base while gathering real-world performance data across different hardware configurations.
8Future development path
The blend model represents just one output from ongoing research into efficient local AI execution. The team identified multiple optimization avenues, including reducing chain-of-thought verbosity (potential 24.5% token savings) and improving verification efficiency. Future releases will incorporate these enhancements while maintaining the balance between reasoning depth and response speed that defines Junie Local's value proposition.
Original article: The JetBrains Blog — Junie Local's Hybrid AI Model Boosts Coding Efficiency