llamaperf

Qwen3.8 27B DaisyRose

on Unknown GPU · oMLX · 262,144 ctx

Tone: positive
Nov 21, 2026
Throughput
50.0 t/s gen
Quant
oQ8e-mtp

Use cases

codingagenticrplong-context

Summary

User compares DaisyRose-27B, an RL fine-tune of Qwen3.8-27B, against the base model on a Mac, reporting generation speed unchanged at about 50 tok/s and memory at roughly 38 GB. Setup is oMLX with the oQ8e-mtp quant, xhigh reasoning and 262k context, identical for both models. On llm-bench.io's four LLM-judged scenarios DaisyRose scores 88.1 overall against 86.9 for base Qwen3.8, a gain the user calls within run-to-run spread; DaisyRose uses about 5% more reasoning tokens, roughly a minute more runtime. The user could not confirm RadixArk's Terminal Bench 4.1 pass@1 claim of 5.6 to 12.0.

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.

Every Qwen3.8 report →