Qwen 3.6 27B is solid up to 262K context. How high have you guys gone above that using Rope/Yarn scaling?

Stack:

• i7 12700K | RTX 3090 TI | 96GB RAM
• Qwen 3.6 27B Q3/Q5 KXL UD

I've been pushing Qwen 3.6 27B above 200K ctx all week, and it handles it like a champ. I'm impressed. Today I hit the ceiling at 262K and it's still functioning well and coherent. I'm planning on trying out Yarn to see if I can push it higher.

How high have you guys pushed it and how has the quality held up?

NOTE: If you're wondering why the context is so high, it's mcp tool and memory/summary bloat (40-50K) on session start, which we're actively working on resolving using an mcp broker server, among a couple of other optimizations.

MODEL SWAPPING: Since a couple of people have asked: I start out with Q5 kv 8/8 until 115K, then I drop to kv 4/4 until 200k, then drop to Q3 kv 4/4 until 262K. Works every time and solid speeds on my RTX 3090 TI.