Breaking

Cognition Unveils SWE-1.7 Coding Model Built on China's Kimi K2.7 Base

July 8, 2026 at 12:18 EDT

Software Dev & Coding
Foundation Models
AI Agents

Cognition on July 8, 2026 announced SWE-1.7, a new model for autonomous coding agents, making it immediately available inside its Devin tooling. Built on China-based Moonshot AI's Kimi K2.7 base model and refined with proprietary reinforcement learning, the company claims it delivers frontier-class performance at a fraction of the cost.

July 8, 2026 · Cognition

SWE-1.7: Frontier-Class Coding at a Fraction of the Cost

Cognition's new autonomous coding model — built on China's Moonshot AI Kimi K2.7 base and refined with proprietary reinforcement learning — closes in on GPT-5.5 while running at ~1,000 tokens/sec and ~one-fifth the price of Opus 4.8.

~1,000

tokens / second via Cerebras

~1/5

the cost of Opus 4.8 per task

$1.97

per task on FrontierCode Main

Benchmarks — SWE-1.7 vs. rivals

Higher is better · columns scaled to score

FrontierCode 1.1 Main

9.4%

30.1%

42.3%

43.0%

46.5%

Terminal-Bench 2.1

39.7%

72.7%

81.5%

84.2%

86.9%

SWE-Bench Multilingual

58.3%

73.5%

77.8%

76.8%

84.4%

SWE-1.6 Kimi K2.7 Code (base) SWE-1.7 GPT-5.5 Opus 4.8

"RL can push capabilities much further than previously believed."

The gain from SWE-1.7 came from an upgraded reinforcement-learning pipeline layered on the Kimi K2.7 base:

Stabilized infra

Reliable RL training pipeline

Data curation

High-quality training signal

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.