Test: pi and Codex beat Claude Code on a local model with 6GB VRAM

A developer ran one local model (LFM2.5-24B-A2B and Qwen3.6-35B-A3B) on a 6GB laptop GPU through three coding agents. pi and Codex solved 12 of 12 tasks, while Claude Code managed 7 of 12, slowed by a ~15,800-token system prompt that made the first request 8.6x longer.

• pi: 12/12 tasks, median bug fix 47 s, prompt ~1,442 tokens
• Codex: 12/12, but LFM made 12 tool-argument errors
• Claude Code: 7/12, prompt ~15,800 tokens, first request 48.8 s
• With Qwen, Claude Code produced no files yet reported success