SWE-1.7 Launch Overview

Cognition has launched SWE-1.7, its most capable model to date, achieving a 42.3% pass rate on the main benchmark, compared to 30.1% for its predecessor, Kimi K2.7.

SWE-1.7 is built on the Kimi K2.7 base model and is designed to optimize performance for longer-horizon asynchronous tasks. It utilizes an extensive data-quality pipeline that ensures each task undergoes automated execution tests, which Cognition claims contributes to training stability and predictable improvements at scale.

In terms of performance metrics, SWE-1.7 achieved an 81.5% pass rate on the Terminal-Bench 2.1 and a 77.8% pass rate on the SWE-Bench Multilingual. Comparatively, Kimi K2.7 scored 72.7% and 73.5% on these benchmarks, respectively. Other models like Code GPT-5.5 and Opus 4.8 also performed better than SWE-1.7 on the main benchmark, with pass rates of 43.0% and 46.5% respectively.

The model is hosted on the Devin platform and provided by Cerebras, capable of processing at 1000 TPS. Inference for reinforcement learning training is supported by Fireworks, while NVIDIA Dynamo manages engine lifecycles and routes inference.

Notably, SWE-1.7 has reduced weight delta transfer size by over 99%, with cross-continental weight updates completed in just 1–2 minutes and minimal inference downtime during updates (3–4 seconds). The maximum rollout duration during training was up to six hours.

This launch follows the introduction of the FrontierCode evaluation framework in June 2026, which aims to enhance agentic software engineering practices. Cognition asserts that SWE-1.7 exhibits significantly different behavior from Kimi K2.7, being more likely to investigate the root cause of bugs, which could lead to more reliable software engineering applications.