Selfhost-GPU TT.22.09.2026: Qwen 3.8 27B on one RTX 5090 for 156K context

Selfhost-GPU TT.22.09.2026: Qwen 3.8 27B on one RTX 5090 for 156K context

| Modell + Quant/Stack | VRAM / GPUs laut Thread | Karte | tok/s oder Fazit | Link |
|—|—|—|—|—|
| Qwen3.8-27B Q6_K, DFlash2 | 30 GB | RTX 5090 | 140-190 tok/s (code), 62 tok/s (prose) | Qwen3.8-27B uncensored Q6_K at 156K context on one RTX 5090, 140-190 tok/s with DFlash2 |

Zusammenfassung

– Modell: Qwen3.8-27B
– Quantisierung: Q6_K
– Software: DFlash2, llama.cpp
– Karte: RTX 5090
– VRAM: 30 GB
– Token Rate: 140-190 tok/s (code), 62 tok/s (prose)
– Zusätzliche Informationen: 156K context, vision projector on the GPU, Sharp chat template. The setup includes a vision projector and a Sharp chat template for enhanced functionality. The model is uncensored and has been optimized for long agentic coding sessions.