Qwen3.6 218K Context on a 3090

On a single RTX 3090, Qwen3.6-27B was pushed further, hitting about 218K context and 50-66 TPS.

Vision input also achieved about 198K context + 51-68 TPS, and a 25K token-scale tool output completed without OOM.

Previously, long tool outputs kept crashing; the cause was that PN12 of the Genesis patch wasn't actually being applied to the code path on vLLM dev205+. apply_all showed success, but the patch wasn't changing anything due to anchor drift, and after fixing this, the tool-prefill OOM disappeared.

The goal isn't maximizing either TPS or context alone, but satisfying all of the following together on a single 3090:

• 200K+ context
• Practically usable throughput
• Stable tool-agent operation

Constraints remain. There's another memory cliff around about 50-60K for single-prompt workloads on a single GPU, which doesn't apply under 2x3090 tensor parallelism. Results depend heavily on quantization and configuration.