I pushed Qwen3.8-27B to 98K context on a 16GB 4080 Super, The quality is surprisingly good!

Article automatically generated from technical news.

⚠️ AI dump incoming: ChatGPT helped me turn several days of testing and increasingly unhinged notes into something readable. The setup, measurements, crashes, configs and tests are all real and were run on my own 4080 Super. Yes, the post is long as hell but two days ago I would have been very happy to stumble across something like this while trying to figure out what a 16GB GPU can realistically run. I've spent the last few days seeing how far I could push Qwen3.8-27B-Escha-W2 on a s