Cognition's SWE-2 achieves 92.8 on Terminal-Bench 2.1
Summary
Cognition's SWE-2 is a 2.8T parameter MoE model with 104B active tokens per, featuring SpecForge and FP8 quantization. It achieves 92.8 on Terminal-Bench 2.1, with pricing notes and deployment details indicating availability on Devin Desktop/CLI and gated API pricing. The article positions SWE-2 among leading models and highlights independent replication as pending.