Marktechpost (@Marktechpost)T2·78 pts
- SWE-bench Pro pass@1 rises from 39.2% to 64.7%; Terminal-Bench 2.1 from 39.5% to 55.8%
- Median turns on solved tasks drop from 59 to 23 with about 4.5x fewer tokens than thinking mode
- Rank-32 LoRA under Apache-2.0, deployable on vLLM or as a llama.cpp GGUF adapter