llamaperf

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

on AMD Strix Halo 128GB · Strata · 24,842 ctx

Tone: mixedAMD hardware
Nov 18, 2026
Throughput
27.6 t/s gen
Quant
IQ3_S (GGUF)
System RAM
128 GB

Use cases

coding

Summary

User benchmarks Qwen3.8-Flash-Next (IQ3_S GGUF, 77.88 GiB) on a 128 GB Strix Halo box, an RTX 5070 Ti and an RTX 5060 Ti, comparing llama.cpp 1.0.0-dev build 12359 with Strata 0.1.59 on a 24,842-token prompt with MTP off. Strata reached 27.60 tok/s on Strix Halo versus 22.16 for llama.cpp, 21.00 versus 5.19 on the 5070 Ti, and 18.70 versus 4.49 on the 5060 Ti. With Strata's MTP enabled, generation rose to 56.7, 25.3 and 19.9 tok/s respectively; native MTP on the 5060 Ti dropped 7.67% on the 24K case. Total response times including prompt processing fell from 52.5 to 33.4 seconds on Halo, 346.0 to 98.8 on the 5070 Ti and 446.8 to 164.3 on the 5060 Ti. Splitting the model across two PCs with llama.cpp RPC was far slower: 0.464 tok/s on a short coding prompt versus 5.776 tok/s on the 5070 Ti alone, 12.45× the decode time per token. The user attributes this to relay traffic and expert-cache misses and asks about direct 10 GbE splits.

Compare this run

Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.

Qwen3.8 125B · 6B active at 4-bit by hardware

Bars are medians on one scale, 0 to 232.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.

Models most run on the AMD Strix Halo 128GB

Bars are medians on one scale, 0 to 77.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.

Everything reported on the AMD Strix Halo 128GB →Every Qwen3.8 report →Estimate Qwen3.8 on this card →