state-m

Researchers achieve 95.3% accuracy on Terminal-Bench 2.1 using DeepSeek V4 Flash and StateM, matching GPT-5.6 Sol Max performance and exploring agent improvement beyond model scaling.