DeepSeek Model Tops Open Coding Benchmark with Auditable Runs

Original: DeepSeek V4 Flash 0731: 82.7% on Terminal-Bench 2.1 with a public harness

Source: antigma.ai ↗