Cognition Releases SWE-1.7 Model Scoring Near Top Frontier Systems
Cognition says its new software-engineering model was trained from a Kimi K2.7 base, runs at 1,000 tokens per second, and keeps gaining from reinforcement learning.
531K13016177.1K
Cognition announces SWE-1.7 which was trained from Kimi K2.7 base performance looks pretty incredible
Scaling01TECH#1218via XStory Brief
Cognition says its new SWE-1.7 announcement on X describes its latest software-engineering model as its most capable yet, with benchmark results that the company says land "within a few points" of top frontier systems while running at 1,000 tokens per second. The company's SWE-1.7 blog post says the model was trained from a Kimi K2.7 base and improved through a refined reinforcement-learning recipe. So far, the core technical, performance, and cost-related claims are coming from Cognition's own launch materials.
Benchmark scores are exciting but more importantly we are seeing incredible results so far using this model in Devin! Try it out:
Scott WuTECH#548
