Artificial intelligence / news brief
MLPerf Training v6.1 adds post-training and recommender tests across 250 results
- Event
- Published

MLCommons released MLPerf Training v6.1 with 250 results from 153 systems submitted by 24 organizations. The suite adds a Qwen3.5-397B reinforcement-learning post-training workload and a DLRMv4-HSTU recommender test. The round includes NVIDIA Rubin, AMD accelerators and Google's TPUv7, with 12 systems using at least 1,000 accelerators.
Why it matters
The new post-training workload gives infrastructure buyers a standardized view of a compute phase increasingly central to reasoning and agent models.
Sources
- MLPerf Training v6.1 Results: Record Participation (external site)MLCommons · primary source ·
- NVIDIA Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1 (external site)NVIDIA · coverage ·
- Crusoe's MLPerf Training v6.1 results on AMD Instinct MI355X (external site)Crusoe · coverage ·
Single source: MLCommons is the official benchmark consortium and publishes the reviewed results; submitter coverage is included for configuration context, not as independent reproduction.