Cognition has announced "SWE-2," a new AI software engineering model based on Kimi K3 with proprietary reinforcement learning (RL). According to the company, SWE-2 utilizes a Mixture-of-Experts (MoE) architecture with 2.8 trillion total parameters and 104 billion active parameters per token.
Cognition Announces Next-Generation SWE-2 Model, Recording High Score on Terminal-Bench 2.1
This article is a translation. Read the Japanese original
In benchmarks, SWE-2 recorded 92.8 on Terminal-Bench 2.1. It also scored 50.0 on FrontierCode 1.1 Main, a figure nearing the 50.9 achieved by Claude Fable 5.1. Conversely, on Terminal-Bench 4.0, which measures long-term autonomous operational capabilities, it remained at 27.3, trailing significantly behind Fable 5.1 (55.8) and GPT-6 Astra (57.9).
SWE-2 is currently available via Devin Desktop and CLI, with deployment to Devin Web and Fusion planned. Cognition has not released the model weights for SWE-2, nor has it disclosed the per-token API pricing at this time.
Sources
- Cognition's SWE-2 achieves 92.8 on Terminal-Bench 2.1 （Hacker News Frontpage, 2026-09-10）
- Cognition Official