Atlas

Models

← All models

SWE-1.7

Cognition · 2026-07-08

Cognition's SWE-1.7 is a proprietary software-engineering model launched on July 8, 2026 after reinforcement-learning post-training from Kimi K2.7 Code. Trained directly in the Devin harness, it uses self-compaction for tasks extending beyond its undisclosed raw context window and produces condensed chain-of-thought; at launch it was available in Devin Web, Desktop, and CLI via Cerebras at 1,000 tokens per second.

Benchmark scores

BenchmarkScore
FrontierCode 1.1 Extended - Pass Rate60.7
FrontierCode 1.1 Extended - Score54.6
FrontierCode 1.1 Main - Pass Rate47.8
FrontierCode 1.1 Main - Score42.3
Loading Atlas data…