The GPU serving record for Kimi K2.6: 511.6 tokens per second.

Brainsless Research Lab now holds the fastest measured GPU serving of Kimi K2.6, the most widely benchmarked open trillion-parameter model: 511.6 tokens per second, single-stream, on four B200 GPUs, lossless. At the record's pin, the fastest provider median on the Artificial Analysis leaderboard read 438.1, and every provider median the board publishes reads below the record. 505.9 was the first GPU reading above 500 on this model.

The record was set at 505.9 on July 5 and raised to 511.6 by blind re-runs of the lab's own public release: gates fixed before the runs, no configuration touched. The configuration uses stock public parts, an open engine (SGLang) and a public draft head, which is why the lab reports the number as a floor. Fovea, the lab's trained drafter, already reads 5.8% above the stock head in a same-session pair.

The result comes out of the paper's real subject: a millisecond-by-millisecond measurement of the trillion-parameter decoding step. The forward pass costs 6.6ms, verifying a drafted token costs 0.4ms, and the rest is engine software, which is where the record came from.

Every number rebuilds from a named artifact, and the whole record re-runs from the repository with one command and about $15 of rented GPU time.