Local Qwen3.8-Flash-Next and Claude Review One Commit
A Reddit user ran Qwen3.8-Flash-Next locally through Claude Code to review one commit in the ChatOKF repository, with Claude Opus serving as the comparison reviewer. The poster reported that Qwen produced eight valid findings with no false positives, including two issues Claude missed, while missing five issues found by Claude and getting two details wrong. The setup used UD-Q4_K_XL quantization, llama-server with ROCm, an AMD Strix Halo 128 GB system, a 262K context window, and an MTP draft head. Reported decode speed fell from about 21 tokens per second at 4K context to about 14 at 128K.
The report presents a first-hand account of using a private, local model as an initial code-review pass while warning that its mechanisms should be verified and it should not be treated as the final word on subtle logic. The single-commit, single-run design and Claude's acknowledged bias as the comparison reviewer make the result an anecdote rather than a benchmark.
I am seriously impressed and want to share my experience.
Qwen3.8-Flash-Next (local) vs Claude in Claude Code: reviewing one real commit
I ran the same code review twice on one commit in my repo https://github.com/leok7v/ChatOKF (Swift, on-device LLM chat app, about 48 files changed, +1048/-102: a module move plus a new 783-line search harness).
Setup: Qwen3.8-Flash-Next UD-Q4_K_XL (unsloth) on llama-server, ROCm, AMD Strix Halo 128 GB, 262K context, MTP draft head, driven headless by Claude Code itself (claude -p pointed at the local server, read-only tools). The comparison reviewer was Claude (Opus) in Claude Code, which reviewed blind first and then judged Qwen's review against the code. Yes, Claude judging its own competitor is a bias; every finding below was checked against the source, and the main one with a small Swift repro.
Result: Qwen produced 8 findings, all real, zero false positives. It caught 2 things Claude missed (a test harness that does not follow the app's recall path, and a markdown ingest that takes the frontmatter "---" as a description). It missed 5 that Claude found, including the most important one: a stop-word filter that compares stemmed words with an unstemmed list, so "this" and "does" leak through as search terms. It also got two details wrong (the exact fallback behavior of a failable init, and column counts off by a few). Severity calibration and the "checked and fine" section were good. It read whole files and chased module boundaries like a careful human.
#Qwen3.8-Flash-Next#Claude#Claude Opus#Claude Code#llama-server#llama.cpp#ROCm#AMD Strix Halo#MTP#ChatOKF