Catching up on a Phoronix story from September 23: Lemonade, the open-source local AI server project aligned with AMD, shipped 2026.39.1 and a 2026.40 release candidate with a couple of notable changes.
What changed
- Streaming models on integrated GPUs now size against the APU’s GTT memory pool instead of the smaller fixed VRAM carve-out. That fixes models like a DeepSeek V4 Flash quant failing on Ryzen AI Max (Strix Halo) machines with plenty of RAM.
- The OpenMOSS ROCm back-end was removed on Windows and Linux because it ran about 40x slower than Vulkan on the same hardware, apparently due to CPU fallbacks.
- A new launch agent for JetBrains’ Junie, and configurable VRAM auto-eviction.
My take
Benchmark your back-ends. “Vendor-native” isn’t automatically faster, and Vulkan keeps proving itself for local inference.
Needs Garlic!
