MLCommons published MLPerf Training v6.1 results on November 16, 2026, bringing the suite’s first large language model post-training measurements into a round covering 153 systems. For university computing teams and AI researchers, the release provides a common test of infrastructure used to improve coding agents, alongside established model-training workloads.
Post-training joins the published results
The consortium’s release reports 250 results from 24 organisations, including the University of Florida. Cloud deployments accounted for 57 systems, or 37%, compared with 29% in v6.0. The largest configuration used 10,240 accelerators. Hardware included NVIDIA, AMD and Google devices, with Rubin-generation GPUs and Ironwood TPUs appearing for the first time.
Two workload changes broaden the suite: Qwen3.5-397B introduces post-training through Group Relative Policy Optimization, or GRPO; DLRMv4-HSTU replaces DLRM-dcnv2 for recommendation, using the Yambda dataset. MLPerf Training co-chair Shriya Rishab said post-training adds architectural complexity and described the submissions as evidence of the community’s rapid uptake of the test.
A coding task that exercises the whole system
MLCommons announced the post-training design on September 24. November’s development is the publication of results. In the workload, a coding agent attempts software repairs inside a sandbox; pass-or-fail rewards from multiple attempts feed GRPO updates, and the updated model generates subsequent attempts.
The selected R2E-Gym subset contains 700 training problems and 251 separate validation problems. OpenHands supplies the agent harness. Evaluation tests are hidden from the agent and protected against modification, preventing it from earning rewards simply by deleting failing tests. The quality target is 0.69 pass@4: solving 69% of validation problems at least once across four attempts.
Training, inference and agent environments compete for CPU time, accelerator resources, memory and storage. That makes the result relevant to research-computing teams allocating resources across a coding-agent experiment. The engineering question extends beyond how quickly an accelerator updates weights: environment execution and generation must also keep pace. For computational reproducibility, the harness and evaluation procedure belong in the experimental record alongside the model.
Google and AMD show different infrastructure lessons
Google Cloud reported ten results across Ironwood TPUs and NVIDIA Grace Blackwell GPUs. Its Qwen post-training submission, result 6.1-0017, used 640 Ironwood chips and reached the benchmark criterion in 18.02 minutes, with a reported score of 0.765 against the 0.69 goal. Google says it ran generation and learning concurrently, keeping training-chip waiting time below 3%.
Google’s timing note defines these measurements as the interval from the official reference checkpoint to the convergence target. The 18.02-minute figure therefore describes a bounded post-training experiment. For a university assessing the result, the checkpoint, model, quality criterion and 640-chip allocation are part of the measurement; the number is not an estimate for building a comparable model from scratch.
AMD’s technical report describes participation in five of the round’s seven benchmarks, covering language models, image generation and recommendation. For Llama 3.1 8B on eight MI355X GPUs, AMD reports a reduction from 86.84 minutes in v6.0 to 72.94 minutes in v6.1, about a 1.19-fold speedup on unchanged accelerator hardware.
AMD attributes the gains to its ROCm and Primus software work. Its account describes combining operations into fewer kernels, reducing intermediate memory transfers, and moving communication onto copy engines to leave compute resources available for training. Those implementation details give computing centres a concrete software question to investigate before assuming that shorter training times require replacement hardware.
The comparison also separates two experimental questions: performance on a fixed machine after software changes, and performance after changing both hardware and software. A university evaluation should preserve that distinction in its evidence. The same-hardware Llama example is useful for a software-upgrade assessment; it does not establish a universal ranking across every workload in the suite.
Read the comparison rules before procurement
The MLPerf Training methodology measures elapsed time to a specified quality target. Results combine repeated trials after removing the fastest and slowest runs and averaging the remainder. Closed-division entries use the reference model; Open-division entries permit changes such as a different model or retraining.
Availability is another filter. Available systems have components that can be purchased or rented; Preview entries must qualify as Available in the next round; Research, Development or Internal entries cover experimental, developing or internal systems. University buyers can use these labels to assess procurement relevance.
Research offices can retain the version, division, accelerator count, software stack and result identifier together. MLCommons links submission metadata and code from its results table. Preserving those details follows the logic of reproducible research practices: another team needs the configuration behind a number to assess or repeat the comparison.
Independent coverage puts software and scale in view
In ComputerBase’s November 17 report, Volker Rißka emphasised software optimisation alongside the new chips and large configurations. His account also highlighted the redesigned results interface and its filters and graphical comparisons. That framing is useful for reading this release: the published performance belongs to a complete configured system, whose software merits attention alongside accelerator identity.
The Latent Daily’s coverage likewise reported the new GRPO workload and the first Rubin and Ironwood training entries. It treated MLPerf Endpoints separately as a test of served inference. For research teams reading announcements that discuss both suites, keeping the benchmark names attached to the figures preserves the distinction between training a model and operating an inference service.
What researchers can examine next
The immediate next step is the published dashboard, where researchers can inspect comparable configurations and their submission details. MLCommons notes that results can be modified or invalidated and maintains a change log. Research offices retaining evidence for a purchasing decision should preserve the result version they used.
For background on the benchmark framework, CASRAI’s MLCommons benchmark definition explains its role. The new release supplies the dated measurements that research teams can now examine against their own training workload and infrastructure requirements.
Sources
- MLCommons: MLCommons Releases MLPerf Training v6.1 Results, Spotlighting Recent Advances in AI Training, checked 2026-11-17
- MLCommons: MLPerf Training, checked 2026-11-17
- MLCommons: MLPerf Training Introduces Its First LLM Post-Training Benchmark, checked 2026-11-17
- ComputerBase: KI-Benchmarks: Nvidia Rubin bis zu doppelt so schnell wie Blackwell, checked 2026-11-17
- The Latent Daily: MLPerf Training v6.1 adds GRPO post-training; Rubin, Ironwood debut, checked 2026-11-17
- AMD ROCm Blogs: Technical Deep Dive into AMD MLPerf Training v6.1 Results, checked 2026-11-17
- Google: New AI training performance benchmarks for Google TPUs, checked 2026-11-17







