Running Qwen3.8 locally on a 128GB laptop for agentic coding: thinking tokens, not tok/s, set the wall clock

deepu105 · reddit · 2026-09-11

The author runs Qwen 3.8 (27B and Flash Next) on an ASUS ROG Flow Z13 (Ryzen AI Max+ 395, 128GB unified memory, Arch Linux) with llama.cpp, his own LlamaStash launcher, and Pi as the coding harness — and argues it can replace Claude Opus 4.6–4.8 for agentic coding if you accept 2–3x longer tasks.

Quality holds up: Flash Next scores 40 on the Artificial Analysis index vs 42 for Opus 4.8; the 27b at xhigh scores 34 vs 32 for Opus 4.6. The 27b one-shotted a whole feature on a large Rust codebase, with Opus 5's review comments mostly nits.
Decode is fine, prefill hurts: 10–15 tok/s decode feels okay, but a cold 31k-token transcript takes 3 minutes to prefill and a full 128k window 18 minutes; warm follow-ups return in 45s.
MTP is the biggest speed win: 7.3 → 22.4 tok/s on an empty window, shrinking to 1.15x at a full 256k.
Flash Next isn't faster per token, it just thinks less: same 5/5 on tasks, 45% fewer tokens, 76.5s vs 289.8s vs the 27b. Thinking is 90–95% of output, so that ratio — not tok/s — sets task duration.

$0/month, fully offline. Full benchmarks, configs and tuning writeup on deepu.tech.

Original post →

Running Qwen3.8 locally on a 128GB laptop for agentic coding: thinking tokens, not tok/s, set the wall clock

More from coding & agent