Qwen3.8 27B Measured on an RTX 5090 — and Why vLLM Didn't Work

Running Qwen3.8 27B on a 32GB card on launch day: the runtime version wall, VRAM measured at each context length, and the full account of trying to move it to vLLM for concurrency and failing. Includes a size survey of the "4-bit" builds — none came in under 18GB — and the image input that a client was quietly discarding.