Local Qwen model drives Claude Code for review: 8/8 real findings, but misses the critical bug

leo-k7v · reddit · 2026-11-03

The author ran Qwen3.8-Flash-Next (UD-Q4KXL, 262K context) locally on an AMD Strix Halo 128GB machine, driven headlessly by Claude Code, to review one Swift commit (+1048/-102) and compared against a blind Claude Opus review.

Results

Qwen produced 8 findings, all real, zero false positives, and caught 2 issues Claude missed
It missed 5, including the most important: a stop-word filter comparing stemmed words against an unstemmed list, letting "this"/"does" leak through
Severity calibration and "checked and fine" sections were solid; it read whole files and chased module boundaries like a careful human

Performance

53 turns, 41 minutes, context grew to 110K tokens
Decode 21 t/s at 4K context, 14 t/s at 128K; prefill 520→230 t/s
MTP gives 1.5x up to 64K but becomes a loss past 100K

Gotchas

Use --lazy-mode or the 27GiB n-gram embedding table loads fully into RAM instead of staying mmapped
unsloth's MTP heads crash llama.cpp (GGMLASSERT); author provides a fix branch
Claude Code sends a mid-conversation system message that Qwen's chat template rejects; override the template

Verdict: a trustworthy local first-pass reviewer for overnight runs, but verify subtle logic yourself; one commit, one run — an anecdote, not a benchmark.

Original post →

Local Qwen model drives Claude Code for review: 8/8 real findings, but misses the critical bug

More from coding & agent