The local development and artificial intelligence tool Junie Local, developed by JetBrains, has recorded a significant performance improvement on Mac devices equipped with M5 Max processors; code generation speed has increased to approximately 70 tokens per second compared to about 40 tokens in the previous version, with the ability to reach burst speeds of up to 107 tokens per second.
This improvement resulted from two major changes introduced to the tool's architecture: the adoption of the new Splash inference engine and the integration of the DFlash2 drafting model, which doubled overall efficiency compared to the previous combination of MTP technology and the MLX-VLM engine.
What has practically changed in the tool's performance?
JetBrains replaced the previous MLX-VLM engine with the Splash engine, developed by Inco AI and specifically supported for Qwen models. This engine allows for the verification of eight drafted tokens at once in each decoding step, while the cost of context attention is significantly reduced to ensure that decoding speed does not decline as the context size grows. Custom int8 operation optimizations for M5 chips have also been integrated into the new engine.
Regarding the drafting model, the tool used the DFlash2 model, which is customized for the Qwen 3.8-27B model. To avoid the overthinking problem that slows down generation, the company relied on the Qwen3.8-3.6-27B-blend version trained using the SpecForge platform on realistic programming paths and code, which increased the acceptance rate and the speed of the coding agent.
Availability and technical requirements
- Speed and generation: Generation speed jumped from ~40 to ~70 tokens per second, and the prefill phase improved to reach 1200 tokens per second compared to 850 previously.
- Memory consumption: The tool requires about 24 GB of RAM when handling a context of 100,000 tokens.
- Supported hardware: Users can now run the tool on Mac computers equipped with M4 processors or newer, provided there is at least 64 GB of RAM, noting that the prefill phase is about four times slower on M4 chips compared to M5.
Users can update the tool via the direct update command in the Junie command-line interface, or perform a full installation using the dedicated script provided by the company, in addition to the availability of DFlash2 models for download via the HuggingFace platform.