Context bottleneck limits local model competitiveness with Claude/Codex

A Reddit user with a 32GB multi-GPU setup reports 27B models run but context remains the bottleneck, making it hard to match Claude/Codex for coding tasks. The discussion highlights trade-offs between local inference and commercial models for developer use.