Benchmarks Desk

The Benchmarks Desk is the collective byline behind the China AI Bench leaderboard (Ranked & Sourced) and the model deep-dives and rankings on this site. It is an organizational byline, not a single person: the desk runs the in-house test harness, keeps the raw logs, and compiles the sourced numbers.

Two models are tested in-house so far, DeepSeek V4 Pro and V4 Flash, run on a 22-task harness in August 2026; task sets, versions, and raw logs are published alongside them. Everything else on the board is compiled: vendor documentation, independent evals (Artificial Analysis, LMArena, NIST CAISI), and developer-reported tests, every number linked to the source we took it from, vendor-reported figures labeled as such. Rows we can't source yet sit in a tracking queue with no scores. We expand in-house testing model by model. When you see "Benchmarks Desk" on a byline, it means the data behind the article is either our raw logs or someone else's that we can point you to, never a number we can't back up.