Local coding models reach practical-use level

Evaluating 27B-32B open-weight coding models on Terminal-Bench 2.0 (89 tasks), Qwen 3.6-27B scored the highest at 38.2% (34/89) under the default per-task timeout.

• The comparison baseline is the default setup, the same as the public leaderboard.
• The author stated that these figures were compared against the verified leaderboard.
• It was noted that MOE models still have more than single-digit room for improvement in inference speed.

The verified SOTA on the public leaderboard is around 80%, and the post interprets the performance of the latest runnable offline models as comparable to hosted frontier models from late 2025.

• Terminus 2 + Claude Opus 4.1: 38.0%
• Terminus 2 + GPT-5.1-Codex: 36.9%
• Claude Code + Sonnet 4.5: 40.1%
• Codex CLI + GPT-5-Codex: 44.3%

The post's conclusion is that coding models runnable in offline/on-premise/regulated environments have, for the first time, come close to the boundary of being practically deployable. The authors summarized this as roughly a 6-8 month performance gap.