Qwen3.8-FN at NVFP4 ( Basalt vs cachy-flash vs ninfer-ext)
Qwen3.8-Flash-Next NVFP4 on one RTX 5090 - prefill and decode against context Full 262,144-token context, one request at a time, MTP speculation on, int8 KV, identical test for all three engines. Every number is tokens/second. Prefill = prompt tokens / time to first token.