Qwen3.8-27B on RTX 5090 Laptop 24GB — TWIN-TURBO NVFP4 (native llama.cpp) vs Coder390 Q3 LynnStyle (.ninfer + NInfer engine in WSL2/Docker): 86 tok/s decode, 262K context, C4 aggregate 228 tok/s. Every number a controlled measurement.
windows docker benchmark mtp wsl2 blackwell kv-cache long-context vision-language-model llama-cpp local-llm gguf speculative-decoding rtx-5090 overthinking twin-turbo nvfp4 24gb qwen3-8 ninfer
-
Updated
Oct 12, 2026 - Python