Qwen3.6 27B NVFP4 + MTP on a Single RTX 5090: 200k Context in vLLM

Qwen3.6 27B NVFP4 + MTP on a single RTX 5090: 200k context working in vLLM

A Reddit user ran Qwen3.6 27B NVFP4 on one RTX 5090 32GB and validated 200k context in vLLM. The setup used fp8_e4m3 KV cache, FlashInfer, and MTP with 3 speculative tokens; a 10-run 200k pass completed with 73.6 tok/s mean generation and 70.2s TTFT. The key constraint is 32GB VRAM: logs showed 8.3GiB KV cache and about 30478MiB total GPU use.

Why it matters: HKR-H/K/R all pass: the hook is single-GPU 200k context, with concrete vLLM settings and 10-run stability data. Reddit sourcing keeps it in the 78–84 band, not P1.