Atlas

Models

← All models

Kimi-Dev-72B

Moonshot AI · 2025-06-16 · 72.7B parameters

Kimi-Dev-72B is Moonshot AI's open-weight 72B coding model built on Qwen2.5-72B-Base and mid-trained on approximately 150B tokens of software-engineering data. It was then cold-started with long chain-of-thought SFT trajectories and optimized by large-scale reinforcement learning on repository tasks, specializing in issue resolution via file localization and code editing with test-suite-verified rewards.

Benchmark scores

BenchmarkScore
EDIT-Bench Complete31.7
EDIT-Bench Complete Easy53.2
EDIT-Bench Complete Hard5.7
EDIT-Bench Core33.3
EDIT-Bench Core Easy54.8
EDIT-Bench Core Hard4.3
SnakeBench Average Apples0.1
SnakeBench Best Apples1.0
SnakeBench Rating14.9
SnakeBench Total Apples2.0
SnakeBench Win Rate33.3
Loading Atlas data…