Unconfirmed: A post on X says a locally run Qwen model found two issues that Claude Opus missed while reviewing a 48-file Swift commit, and Automatica has not confirmed the claim.
@EugeneSmarts wrote on 3 November1 that the Qwen model made eight findings, none of which Claude Opus judged to be a false positive. The post says Qwen also missed five findings that Opus caught. The poster does not claim the local model beat Opus overall, and the results he reports are mixed.
According to the post, Opus reviewed the commit blind first and then judged Qwen's findings. The "zero false positives" figure is therefore Opus's assessment, not an independent one. The post names two issues that Qwen found and Opus did not: a test harness that did not follow the app's recall path, and a markdown ingest routine that treated frontmatter as a description. The most notable miss it lists is a stop-word filter bug in which stemmed words were compared against an unstemmed list.
The post describes the setup as "Qwen3.8-Flash-Next UD-Q4_K_XL unsloth" on llama-server with ROCm. It ran on an AMD Strix Halo machine with 128GB of memory, 262K context and an MTP draft head, driven headless by Claude Code. The poster reports 21 tokens per second at 4K context, falling to 14 at 128K, over 53 turns and 41 minutes. He also lists several workarounds, including a lazy-mode flag for the model's 27 GiB embedding table and a patch for a llama.cpp crash with Unsloth MTP heads.
The poster's verdict was: "A trustworthy private first-pass reviewer for zero cost, but always verify subtle logic manually before shipping."
The "unsloth" label in the setup line suggests the quantised build came from Unsloth, which could help trace the model. The post links no model repository, though, and Automatica could not find a public model card or weights for "Qwen3.8-Flash-Next" in the evidence supplied. The commit and the full findings list are not public either, and the only judge named is Claude Opus. The poster's account has about 4,200 followers, and the post had 11 likes when Automatica observed it. Neither figure bears on whether the claim is accurate.
The post does not say who owns the Swift project. Automatica found no statement from Alibaba's Qwen team, Unsloth or Anthropic about the post or the model name, and has no earlier reporting on this claim.
Three things would confirm or undercut it: a public model card or weights for Qwen3.8-Flash-Next, publication of the commit and the findings list, and an assessment by someone other than Claude Opus. A repository that does not match the post's description would undercut it.