AI-generated editorial illustration; not a documentary photograph.
JetBrains has updated Junie Local with Qwen3.8-3.6-27B-blend, a 27B coding model created by combining Qwen3.8-27B and Qwen3.6-27B weights in equal proportions. Apple M5 users can install the model through Junie, while a nightly build introduces experimental Windows support for compatible NVIDIA RTX cards.
A middle ground on reasoning cost
The new model is intended to narrow the gap between Qwen3.6, which JetBrains previously selected for laptop use with reasoning disabled, and the more capable but slower Qwen3.8. JetBrains says the blend requires no additional post-training because its two source models share an architecture.
In the company's 100-task internal coding benchmark, Qwen3.8-3.6-27B-blend completed 37 tasks. Qwen3.6 completed 34 without reasoning, while Qwen3.8 solved 39 but generated substantially more output. JetBrains measured 71% fewer output tokens for the blend than Qwen3.8. Among 30 tasks both models solved, the totals were 279K and 935K tokens respectively.
Results vary by workload
JetBrains also reports that four LiveCodeBench runs produced an average score of 85.47% for the blend, against 83.29% for Qwen3.8 at a similar output cost. Qwen3.6 averaged 67.87% across four complete passes and consumed about 24.1 million output tokens per pass, compared with approximately 6.14 million for the blended model.
Those gains did not carry uniformly into visual benchmarks. The blend generated fewer tokens than Qwen3.6 with thinking enabled but more than Qwen3.8, which JetBrains attributed largely to longer reasoning. The company also warns that the model can repeatedly revisit an unsuccessful approach, requiring users to interrupt it and restart with a narrower objective.
Runtime limits and Windows preview
Junie Local uses multi-token prediction to draft several tokens for parallel verification. On an M5 MacBook Pro, proposing two tokens per round improved decoding speed by 60% over disabling the feature, according to JetBrains; proposing four reduced the gain to 36%. The company retained a four-bit MTP head because an eight-bit version showed no consistent speed advantage and used more memory.
Long contexts remain a constraint because attention verification became slower as context grew, despite proposal accuracy holding up. On Windows, the experimental nightly release supports NVIDIA RTX cards using Ampere or newer architectures with at least 24 GB of VRAM. The model is also available through Hugging Face.
Source: Making Local AI Smarter and Faster
SIGNAL TITANS
Leave a Reply