Terminal-Bench 2.0 · Local Models · n=89

The 89-task terminal-agent scoreboard

Every local model we ran through Terminal-Bench 2.0 — a multi-turn shell-agent benchmark — task by task. Who solved what, where the frontier of local capability actually sits, and how a 2-bit MiniMax-M2.5 holds up against the Qwen and Gemma field.

Harness Harbor + terminus-2 Env Docker linux/amd64 Concurrency 1 Budget --agent-timeout-multiplier 0.5 Measured on-rig 2026-05 → 07

Ranking

pass-rate, n=89

✳ MiniMax IQ2_M plotted at its fair 0.5×-budget score (25.8%). With a full/1.5× agent-time budget it recovers to ~30.3% (dashed marker) — a diagnostic, not cohort-comparable, since only it got the extra time. The rest of its gap to the frontier is structural (2-bit throughput + tool-call format instability).

The shape of the field

Per-question matrix

89 tasks × 9 models
MiniMax passed Budget-recoverable MiniMax-only pass Nobody passed Solved by ≥1
✓ pass ✕ ran, failed verifier ⧗ agent timed out · infra error / no data + MiniMax recovered at higher budget