Community · Nov 2, 2026, 13:08 UTC

Local Qwen3.8 models score near Claude Haiku 5.5 on nine real C++/CUDA bug fixes

A Reddit benchmark of nine private regressions from a local inference engine scored Claude Opus 5.5 at 0.74 and Qwen3.8-Flash-Next at about 0.5, with 3-bit quantization costing tokens rather than score. Results carry wide confidence intervals of about ±0.2 on nine tasks.

• LabAnthropic
• LabAlibaba
• ModelClaude Opus 5.5
• ModelClaude Sonnet 5.5
• ModelClaude Haiku 5.5
• ModelClaude Sonnet 5
• ModelClaude Sonnet 4.5
• ModelClaude Haiku 4.5
• ModelQwen3.8-Flash-Next
• ModelQwen3.8-27B
• ModelGemma 4 26B-A4B
• Topicbenchmark
• Topiccoding
• Topicopen-weights