Loading post…
JetbrainsVladimir SotnikovAnnouncement
Making Local AI Smarter and Faster
Summary
Junie Local’s update introduces Qwen3.8-3.6-27B-blend, a merged model that balances speed and capability for local coding agents. The blend, created by merging Qwen3.6 and Qwen3.8 weights in equal proportions, completed 37 tasks in a 100-task internal benchmark, compared to 34 for Qwen3.6 and 39 for Qwen3.8, while generating 71% fewer output tokens than Qwen3.8. The runtime uses multi-token prediction (MTP), with a two-token proposal on M5 MacBook Pro making decoding 60% faster. A four-bit MTP head was kept over an eight-bit one due to similar acceptance rates and memory savings. The update also brings experimental NVIDIA RTX support on Windows for machines with at least 24 GB of VRAM.
Context
The post addresses the trade-off between speed and capability in local coding agents, where earlier versions of Junie Local forced users to choose between a faster model (Qwen3.6) and a more capable but slower one (Qwen3.8 with reasoning enabled).
Approach / What changed
The approach is to merge the weights of Qwen3.8-27B and Qwen3.6-27B in equal proportions to create a single 27B model, Qwen3.8-3.6-27B-blend, without additional post-training. The runtime uses multi-token prediction (MTP) with a four-bit head to improve decoding speed.
Takeaways
The Qwen3.8-3.6-27B-blend model completed 37 tasks in a 100-task coding benchmark, compared to 34 for Qwen3.6 and 39 for Qwen3.8, while generating 71% fewer output tokens than Qwen3.8.
Using a two-token MTP proposal on an M5 MacBook Pro made decoding 60% faster than running without MTP, but increasing to four tokens reduced the speedup to 36% due to extra GPU work.
A four-bit MTP head accepted 63.0% of proposals, nearly matching the 63.6% acceptance rate of an eight-bit head, with no consistent speed advantage, so the four-bit version was kept to save memory.
FAQ
How was the blend model created?
The blend model was created by merging the weights of Qwen3.8-27B and Qwen3.6-27B in equal proportions. Since both models share the same architecture, this produces a single 27B model without any additional post-training.
Why did the blend model use fewer tokens than Qwen3.8?
The blend model generated 71% fewer output tokens than Qwen3.8. On the 30 tasks completed by both models, the blend used about 279K tokens versus 935K for Qwen3.8, and it used fewer tokens on 29 of those 30 tasks, confirming its token efficiency.
What hardware is supported for the experimental Windows version?
The experimental Windows support covers all NVIDIA RTX cards based on Ampere or newer architectures with at least 24 GB of VRAM. It is available via the nightly build of Junie.
Source and freshness
- Source
- blog.jetbrains.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.