Where we watch intelligence move.

We don't rank frozen models. We build the instruments to observe a system improving itself — before the loop, after the loop, and the distance between.

SWE-Prometheus

Repository-level engineering governance under real-world constraints.

Top 5NGI mean
  1. 1MoonshotAIKimi-K30.5760
  2. 2ZhipuGLM-5.3-Flash0.5760
  3. 3AnthropicClaude-Opus-50.5293
  4. 4AlibabaCloudQwen3.8-Max0.4630
  5. 5ZhipuGLM-5.20.4438

SWE-PolyVision

Cross-image abductive reasoning for repository-level software engineering.

Top 5Best E2E
  1. 1MoonshotAIKimi-K350.0
  2. 2ZhipuGLM-5.3-Flash39.6
  3. 3OpenAIGPT-5.6-Sol37.5
  4. 4DeepSeekDeepSeek-V4-Flash31.2
  5. 5AlibabaCloudQwen3.8-Max29.2

ChemArgus

Fine-grained, multimodal, diagnostic evaluation of competition-level chemistry reasoning.

PDFGitHubSoonHugging FaceSoon
Top 5Full problem %
  1. 1AnthropicClaude Opus 561.5
  2. 2ZhipuGLM-5.252.0
  3. 3GoogleGemini 3.8 Flash48.5
  4. 4DeepSeekDeepSeek V4 Pro36.7
  5. 5AlibabaCloudQwen3.8-Max36.0