LM Studio Qwen 3.6 Crawls at 4 t/s

The hot take: this looks less like a broken app and more like a capacity mismatch plus backend selection issue. On this class of hardware, a 27B dense GGUF at Q4_K_M is already on the edge, so once context and runtime overhead kick in, throughput can fall off a cliff.

• –The Reddit thread’s consensus is blunt: 27B dense models do not offload cleanly on 16GB VRAM, so partial system-memory execution is expected to be slow.
• –The model itself is roughly 16.5GB in Q4_K_M form, which leaves essentially no headroom for KV cache, context, and runtime buffers on a 16GB card.
• –LM Studio on Linux can use llama.cpp backends, but if it is picking the wrong backend or prioritizing the iGPU first, performance can degrade sharply; the fix is to verify CUDA/Vulkan selection and confirm the dGPU is actually active.
• –The best practical workaround is to drop to a smaller quant or switch to a more GPU-efficient architecture like a 35B A3B/MoE variant, which the commenters suggest will run much better on this machine.
• –The user’s CPU hitting 170% while RAM and VRAM stay around 8GB is a strong sign that inference is not staying on the GPU in the way they expect.