A Shared AI Benchmark Now Tests the Tuning That Agents Rely On

MLCommons published results for its MLPerf Training v6.1 benchmark suite on Nov. 16, 2026, reporting 153 submitted systems, 250 test results and 24 submitting organizations. The round adds the first post-training benchmark the training suite has carried.
Post-training is the work done on a model after its first full training run: tuning a model that already exists on particular tasks, rather than building its general ability from scratch. The release says much of the recent innovation in AI, including software that carries out tasks on its own, depends on post-training, and that the MLPerf community designed and built the new test in response.
That test, Qwen3.5-397B, post-trains a model to reason its way through software engineering tasks, including calling outside tools. The round's other new benchmark, DLRMv4-HSTU, measures recommendation systems and replaces an earlier recommender test. MLCommons describes the suite as an open-source, peer-reviewed set of whole-system tests covering models, software and hardware.
The largest system in the round used 10,240 accelerator chips, according to MLCommons, and the hardware submitted came from NVIDIA, AMD and Google. The round carries the first published results for NVIDIA's Rubin-generation GPUs and for Google's TPUv7, also called Ironwood.
"Post-training has added a new level of complexity to the architecture of AI training," said Shriya Rishab, a co-chair of the MLPerf Training working group.
NVIDIA said in a blog post the same day that it had swept every benchmark in the round. That is the company's own account of its own submission; MLCommons' release reports the round's totals and names no top performer.
Among the submitting organizations was one newcomer. "We would especially like to welcome first-time MLPerf Training submitter Crusoe," said David Kanter, head of MLPerf at MLCommons.
