---
title: "MLPerf Training v6.1: 153 Systems, 250 Results and Two New Benchmarks, Including Qwen3.5-397B"
url: https://aibenchmark.gr/en/news/mlperf-training-v6-1-153-systems-250-results-two-new-benchmarks-qwen3-5-397b/
language: en
published: 2026-11-16T16:19:25.354Z
updated: 2026-11-16T17:19:30.010Z
site: "AI Benchmark"
section: "Benchmarks"
---

# MLPerf Training v6.1: 153 Systems, 250 Results and Two New Benchmarks, Including Qwen3.5-397B

On 16 November 2026, MLCommons announced the MLPerf Training v6.1 results, with 153 systems and 250 results, plus DLRMv4-HSTU and Qwen3.5-397B as new benchmarks.

> Written with AI from the sources listed below. Editorial responsibility: Konstantinos Perros. How we write: https://aibenchmark.gr/en/how-we-write/

## Key facts

- The MLPerf Training v6.1 round drew 153 submitted systems, which produced 250 results, according to MLCommons. [1]
- Two benchmarks were added: DLRMv4-HSTU for recommender systems and Qwen3.5-397B for large language model post-training. [1]
- 57 of the 153 systems, or 37%, were in the cloud, compared with 29% in v6.0. [1]
- Twelve systems used 1,000 or more accelerators, and the largest reached 10,240. [1]
- NVIDIA Rubin-generation GPUs and the Google TPUv7 (Ironwood) appeared for the first time. [1]

## The two new benchmarks in v6.1

MLCommons, an open AI engineering consortium, announced on 16 November 2026 the results of version 6.1 of the MLPerf Training suite. The suite consists of full-system tests that stress models, software and hardware across a range of machine learning applications. It is open source, peer-reviewed, and its collection of benchmarks is curated by a group of AI community experts.

The first new benchmark, DLRMv4-HSTU, targets recommender systems and replaces the DLRM-dcnv2 of earlier editions. It uses the attention-based HSTU architecture and is trained on the Yambda dataset, which MLCommons describes as one of the largest open recommendation datasets.

The second, Qwen3.5-397B, is, according to MLCommons, the first of its kind in MLPerf Training: large-scale post-training with Group Relative Policy Optimization (GRPO). The workload trains the model to solve software engineering tasks, with specialized duties such as tool-calling. The organization explains that many recent advances, such as agentic capabilities, rely on post-training activities, and that is why the MLPerf community designed this particular test.

## Scale and participation: the round's figures

According to MLCommons, the round surpassed some of the records set by v6.0, which was published six months earlier, in number of submissions, systems, results and typical scale. 153 systems and 250 results were submitted. Of the systems, 57 (37%) were cloud, whereas in v6.0 the share was 29%.

Twelve systems had 1,000 or more accelerators, with the largest reaching 10,240. The hardware came from NVIDIA, AMD and Google. NVIDIA Rubin-generation GPUs and the Google TPUv7 (Ironwood) made their debut.

24 organizations took part in the round: AMD, ASUSTeK, Azure, Cisco, CoreWeave, Crusoe, Dell, Firmus, Fujitsu, Giga Computing, Google, HPE, Inventec, KRAI, Lambda, MiTAC, Nebius, NVIDIA, Oracle, Quanta Cloud Technology, Supermicro, tiny corp, TTA and the University of Florida. Crusoe submitted to MLPerf Training for the first time.

## What the organizers say and how to read the results

Pavan Yalamanchili, co-chair of the MLPerf Training working group, said that greater participation and a wider variety of configuration choices give a fuller picture of how systems compare with one another in realistic training scenarios. He added that as tasks become more specialized, the hardware being deployed changes too. According to him, the data helps an organization find the right size for its own deployment, from a single-node system to a multi-node cluster with thousands of accelerators.

Shriya Rishab, also a co-chair, said that post-training has added complexity to the architecture of AI training. The new benchmark is the result of two years of work and, in her words, the results show that the community is adopting it quickly. David Kanter, head of MLPerf at MLCommons, described the participation as a sign of the value of trustworthy data for new workloads and new systems.

MLCommons says the Training benchmark page contains the full results and that the release includes new visualization tools. With these, one can compare performance per system and rank results of systems of similar scale. Per-submitter information is in the supplementary material. The organization also invites submitters and other interested parties to join the working group.

## Questions and answers

**Which benchmarks were added to MLPerf Training v6.1?**

DLRMv4-HSTU, which replaces DLRM-dcnv2 and is trained on the Yambda dataset, and Qwen3.5-397B, a post-training workload using GRPO, were added.

**How many systems and organizations took part?**

153 systems with 250 results were submitted by 24 organizations. 37% of the systems were cloud.

## Sources

1. MLCommons: [MLCommons Releases MLPerf Training v6.1 Results, Spotlighting Recent Advances in AI Training](https://mlcommons.org/2026/11/mlperf-training-v6-1-results)
