LLM Reference

SWE-2 Models by Cognition

CognitionProprietaryCodingAgents
1 model2026

Last refreshed 2026-09-10. Next refresh: weekly.

Details

ResearcherCognition
LicenseProprietary
Commercial useCommercial use: conditional
Models1
Released2026

Capabilities

ReasoningAll models
JSON / Tool useAll models
Code ExecutionAll models

Links

Website

About

Cognition's SWE-2 coding model family for agentic software-engineering work, available through Devin product surfaces and optimized for cost-efficient frontier coding performance. Launched September 10, 2026, post-trained from the Kimi K3 2.8T-parameter base with multi-trillion-parameter RL that trains all reasoning-effort levels in a single run.

Decision facts

Best fit
codingagentsreasoning
Capability starting point
SWE-2 with reasoning and JSON / Tool use
Lowest tracked input
Not tracked
Closest related family
Agents-A1

Current Variants

Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.

1 in view
SWE-2Current

Use when the workload needs coding, reasoning, and JSON / Tool use.

2026-09codingreasoningJSON / Tool use

Release Timeline

1 release group
2026-09
1 current
SWE-2
codingreasoningJSON / Tool use
Current

Specifications(1 models)

SWE-2 model specifications comparison
ModelReleasedParametersReasoningJSON / Tool useCode Exec
SWE-22026-092.8T (base: Kimi K3)YesYesYes

Available From(1 provider)