llamaperf

Qwen3.8 125B (6B active) Flash-Next

on NVIDIA RTX 5080 · Basalt · 65,536 ctx

Tone: positiveNVIDIA hardware
Nov 20, 2026
Throughput
193.2 t/s gen · 4589.0 tokens/s prompt processing (prefill)Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.
Quant
IQ2_XS (GGUF)
KV cache
int8
System RAM
64 GB

Use cases

coding

Summary

User benchmarks Qwen3.8-Flash-Next at 193.2 t/s decode and 4,589 tok/s prefill on an RTX 5080 with Basalt 0.5.0, against 155.6 t/s and 3,785 tok/s on Strata 0.1.55. Setup is IQ2_XS weights with a drafter, a 65,536-token prompt and an int8 KV cache, median of 3 runs with no prompt cache reuse and a 1,024-token output cap. At 120,071 tokens Basalt reached 205.6 t/s decode and 4,661 tok/s prefill versus Strata's 175.9 t/s and 3,706 tok/s; Basalt finished 18-20% sooner overall. The user is switching their daily engine to Basalt.