MLPerf Training now measures more of the work that turns a base model into a useful agent. MLCommons released Training v6.1 on 16 November 2026, adding a large-scale reinforcement learning post-training workload alongside a new recommendation benchmark. For infrastructure buyers, that expansion is more consequential than another chart declaring a winning accelerator.
The new tests cover different work, while vendor announcements combine results from different benchmark families. A training time, a per-GPU comparison and an inference throughput-per-watt figure answer different purchasing questions. They cannot be substituted for one another.
Post-training joins the measured workload
The new Qwen3.5-397B benchmark uses Group Relative Policy Optimization, or GRPO, for post-training on software-engineering tasks. MLCommons describes it as the suite's first large-scale post-training workload. The second addition, DLRMv4-HSTU, replaces the earlier DLRM-dcnv2 recommendation test and uses an attention-based architecture with the Yambda dataset.
That replacement matters for historical charts. A faster result on a newly defined recommendation task does not establish that an unchanged older task became faster. The model, dataset and quality requirement belong to the comparison, not just the benchmark's family name.
MLCommons also reports first appearances for NVIDIA Rubin-generation GPUs and Google Ironwood TPUs in this round. Their inclusion broadens the published evidence, but the hardware name is still only one part of a submitted system.
The post-training system includes more than the update step
The MLPerf Training rules define a benchmark as learning to a target quality level and a run result as the required wall-clock time. For the new post-training workload, the rules constrain the agent harness and the task contract. Dedicated CPU capacity used for its sandbox counts toward the submission's overall scale, even when that capacity runs separately from training and generation.
This changes the infrastructure question. Capacity that updates model weights is not the only capacity involved in an agent learning from attempts. The system also needs to generate those attempts and support the environment in which they are evaluated. Buying more accelerator compute for one stage does not, by itself, demonstrate that the complete loop will finish sooner.
That is an architectural implication of the workload, not a claim that MLPerf has measured a bottleneck in every customer's deployment. A production workflow with different tools, task durations or sandbox resources can behave differently from the reference task.
AMD's broader coverage is evidence about particular workloads
In its 16 November technical account, AMD says it submitted to five of the seven benchmarks, using Instinct MI355X and MI350X accelerators for its first-party results. It identifies language-model pretraining and fine-tuning, recommendation and image generation among the covered work.
AMD also describes ROCm and Primus optimizations on existing hardware, as well as comparisons involving a change of hardware generation. Those are different explanations for improvement. A same-hardware software gain supports the value of a better implementation; a comparison that changes both hardware and software cannot isolate the contribution of the chip alone.
Coverage should remain attached to the individual workload. Absence of a submission is not a measured failure, and presence on five tasks is not evidence of performance on the remaining two. A buyer planning post-training needs results for that task, rather than assuming that a strong pretraining entry answers it.
NVIDIA's inference footnote belongs to a different comparison
NVIDIA's announcement discusses both Training v6.1 and MLPerf Endpoints. Its Endpoints example compares Kimi K3 throughput per watt under a specified interactivity condition. The accompanying footnote labels the Vera Rubin result provisional and pending MLCommons verification, while identifying the compared GB300 result as verified.
That status qualifies the Endpoints example. It should neither disappear when the claim is repeated nor be reassigned to unrelated Training results. Likewise, an inference efficiency claim does not measure training energy, and a time-to-train result does not establish production inference latency.
Our MLPerf Inference explainer covers serving scenarios and latency constraints. This release adds a different decision: which complete training system reaches the required learning target, including the new post-training task.
Faster completion still needs a resource denominator
A training result records elapsed time for its submitted configuration. It does not automatically include the commercial price of renting that configuration or the cost of recovering a failed production run. Those remain procurement inputs, as discussed in our training and inference economics analysis.
For a specific shortlist, the meaningful evidence is the workload, quality target, accelerator and CPU allocation, software revision and measured completion time for each candidate. Comparing entries at different scales can illuminate a deadline-versus-resource trade-off. Dividing a result by device count cannot turn a distributed run into a measurement of a single device.
The v6.1 addition is therefore a reason to revise evaluation plans. A team that previously shortlisted hardware using only pretraining results now has a standardized post-training task to examine. The next question is whether its own agent loop resembles that task closely enough for the result to inform the purchase.

