Skip to content

Fixed-five SWE results

All rows use the same five-task denominator. A qualifying result needs complete official grading, at least four resolved tasks, and both Django and Pylint sentinels resolved. Every row below has sentinels=false; no row authorizes a promotion. complete describes grading completeness, not qualification.

Lane Attempted Resolved / 5 Graded / 5 Empty Complete Sentinels Native record
GLM official-base baseline 5 2 5 0 true false native
Coder390 INT4 TP1 initial 5 3 3 2 false false native
Coder390 INT4 TP1 medium 5 3 5 0 true false native
Coder390 INT8 TP1 medium 5 2 5 0 true false native
GLM mixed TP2 no-spec 5 1 1 4 false false native
GLM mixed TP2 thinking-off 5 1 2 3 false false native
Qwen QAD TP2 xhigh 5 1 2 3 false false native
DeepSeek TP1 low 5 2 4 1 false false native
DeepSeek TP1 max 5 1 1 4 false false native
Qwen EXL3 TP1 xhigh 5 4 4 1 false false native
Qwen EXL3 TP1 medium 5 3 5 0 true false native
Stock Qwen FP8 TP1 xhigh 5 2 4 1 false false native
Ignis R4 terminal 5 2 2 3 false false native

The Ignis result resolved Django; Pylint was ungraded with an empty patch, so the required sentinel pair remains unproven and sentinels=false. The table preserves each result rather than ranking unlike configurations.