What this is not
Not a mirror of LMSYS or Artificial Analysis. Those boards move daily and several vendors publish their own evals on the same name. A score without a who-ran-it is advertising.
Who ran the score.
8 scores on model cards. 0 independent. 8 vendor-published. 0 unclear. No invented leaderboard.
Independent layer is an honest zero. LiveBench, LMSYS, HELM, SWE-bench, Aider — only when a card cites them. We do not scrape a board to fill this.
Vendor
DeepSeek V4 Flash Vision Exp · DeepSWEVendor · 59.3 · DeepSeek official changelogDeepSeek V4 Flash Vision Exp · Terminal-Bench 2.1Vendor · 83.9 · DeepSeek official changelogDeepSeek V4 Pro · Terminal-Bench 2.1Vendor · 87.9 · DeepSeek official changelogGemini 3.1 Pro Preview · SWE-bench VerifiedVendor · 80.6% · Google model cardGLM 5.2 · SWE-bench ProVendor · 62.1 · official docs.z.aiGLM 5.2 · Terminal-Bench 2.1Vendor · 81.0 · official docs.z.aiGLM 5.3 · DeepSWE v1.1Vendor · 66.9 · vendor chart (Z.ai), no independent audit yetGLM 5.3 · Terminal-Bench 3.0Vendor · 28.3 · vendor chart (Z.ai), no independent audit yet