SWE 2
SWE-2 is Cognition's agentic coding model built for autonomous software engineering, post-trained from a 2.8T-parameter base model using reinforcement learning. It runs at three effort levels (medium, high, max) and is designed for focused codebase exploration, strong test coverage, and resourceful problem-solving when standard fixes are blocked.
- Type
- Coding
- Weights
- Closed
More from Cognition
10
-
Devin 1
Text
Devin 1 is Cognition’s AI software engineer that plans and executes coding tasks end to end inside a sandboxe…
-
Devin 1.1
Text
Devin 1.1 is a Devin update that adds full API functionality and improves speed, cost-efficiency, and customi…
-
Devin 1.2
Text
Devin 1.2 is a Devin update focused on better reasoning with repository context, plus new collaboration and a…
-
Devin 2.1
Text
Devin 2.1 adds confidence ratings for tasks, stronger built-in codebase intelligence for Q&A and follow-ups,…
-
Devin 2.2
Text
Devin 2.2 is a major Devin update that adds stronger autonomy via computer use for end-to-end testing, plus s…
-
SWE 1.5
Multimodal
SWE-1.5 is Cognition’s fast software-engineering agent model designed for strong coding performance with very…
-
SWE 1.6
Text
SWE-1.6 is Cognition’s latest model for software engineering agents, built to improve both raw capability and…
-
SWE 1.7
Coding
SWE 1.7 is Cognition's most capable coding model to date, reaching frontier level intelligence at a much lowe…
About Cognition
1
Cognition develops autonomous AI software-engineering agents, including Devin.