Junie Local became 70% faster on Mac M5 - The JetBrains Blog
Blog post from JetBrains
JetBrains reports that Junie Local for Mac M5 has increased coding-generation performance from roughly 40 to 70 tokens per second, with bursts reaching 107 tokens per second, through adoption of the Splash inference engine and newly fine-tuned DFlash2 draft models. Splash improves speculative decoding by verifying multiple drafted tokens per step and reducing the context-related cost of attention, while JetBrains also ported its M5 int8 optimization from MLX-VLM to the new engine. To reduce excessive reasoning-token generation associated with Qwen 3.8, the update uses the Qwen3.8-3.6-27B-blend model and a DFlash2 drafter fine-tuned on internally generated coding-agent trajectories, achieving strong acceptance rates with about one million generated training tokens. The release also raises prefill speed from 850 to 1,200 tokens per second, reduces memory use to about 24 GB for a 100,000-token context, and expands support to M4 and newer Macs with 64 GB of RAM, although M4 prefill remains substantially slower than M5.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| MLX | 5 | 55 | 14 | 7 | -30% |
| AI Agents | 2 | 7,016 | 1,668 | 272 | +24% |
| AI Coding Assistant | 2 | 1,961 | 544 | 148 | +23% |
| LLM | 1 | 6,578 | 1,243 | 236 | +9% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.