Benchmarks show Qwen3.8-Flash-Next IQ3_S running on Basalt with an RTX 5080 (16GB VRAM) and 62GB DDR5-6800 RAM achieves up to 5,662 tokens/s prompt processing and 262 tokens/s generation across 262K context windows with int8 KV caching. In agent coding sessions with 103 requests up to 130K context, the model sustained 5,011 tokens/s prompt processing and 153 tokens/s generation while using 15.7GB VRAM and 54.1GB system RAM. Memory usage remained stable during context sweeps from 15K to 207K tokens.