I swapped it in just a moment ago and asked Claude to write a comparison. I haven't tried it myself yet, so I can't say how it feels. Below is what Claude wrote.
Qwen3.8 27B — you can just swap it in over 3.6 (RTX 3090)
I switched to the unsloth GGUF Q4. With the settings unchanged, it runs fine without rebuilding llama.cpp.
Before downloading I compared the two models' config files, and the internal structure is identical to 3.6. So it loads right up on the build you were already using.
Measured results (3090 24GB, same question and same settings, MTP on for both)
3.6 3.8
Speed 47.6 tok/s 47.5 tok/s
Memory 23.3GB 22.7GB
MTP (the speed-acceleration feature built into the model) stays on and works in 3.8 as well. I even checked the server status. The speed is the same, and 3.8 uses about 0.5GB less memory.
Image recognition, Korean (no Chinese characters mixed in), and tool calls were all fine.
I did not compare quality. I only measured with a single short question. All I confirmed is that "swapping it in does not make things slower or broken."
One tip: don't delete your existing model files. Rolling back takes 30 seconds.
