What was added and what was retired in round v6.1
MLCommons' MLPerf Training v6.1 round, with verified Closed division results published on 16 November 2026, includes 250 submissions from 25 organizations. In v6.0 there were 229 submissions from 24. Two benchmarks were added. The first is Qwen3.5-397B GRPO, an agentic reinforcement learning workload in which a 397-billion-parameter model learns to repair real software projects. The second is DLRMv4-HSTU, a generative recommendation model. Seven organizations ran the first with 29 submissions and eight ran the second with 20, and MLCommons notes that first-round numbers partly reflect the novelty factor.
Two benchmarks were retired. Llama 3.1 405B, a dense pretraining model, left after three rounds, because MLCommons considers that large dense models no longer represent large-scale pretraining, which has shifted to mixture-of-experts models. DeepSeek-V3 and Qwen3.5-397B fill that role. DLRM-DCNv2 was retired after seven rounds and replaced by DLRMv4 (HSTU).
Scale increased. The largest Closed system reached 10,240 GPUs, up from 8,192 in v6.0, with NVIDIA's GB300 NVL72 and Azure's GB200 NVL72 training DeepSeek-V3. Of the systems, 76% (117 of 153) used two or more nodes, compared with 61% (72 of 118) in v6.0. The median number of accelerators per Closed result was 8 in every round from v4.0 to v6.0 and is now 32. The round features 12 accelerator products, with Google's Ironwood and NVIDIA's Rubin making their debut as Available systems.
MLCommons also provides a progress index, which it describes as unofficial and derived from the results. Across the four benchmarks that carried over unchanged from v6.0 to v6.1, the geometric mean of the gains in the fastest result is 1.49. The gains are uneven: FLUX.1 improved markedly in about half a year, while the smaller Llama 3.1 8B improved much less.
Vera Rubin NVL72 results according to NVIDIA
The figures below come from NVIDIA's technical blog, that is, from the manufacturer about its own hardware, and not from an independent evaluation. The company says that in its debut the Vera Rubin NVL72 achieved the fastest training time on six of the seven benchmarks, with up to 2x the performance of an optimized GB300 NVL72. On the seventh, DeepSeek-V3 671B, the absolute fastest time of the round came from a cluster of 10,240 GB300 NVL72 GPUs, at 1.38 minutes. NVIDIA adds that its platforms were the only ones to submit all seven benchmarks.
On DeepSeek-V3, a 671-billion-parameter mixture-of-experts model, the Vera Rubin NVL72 trained in 4.14 minutes with 1,024 GPUs across 16 racks. The GB300 NVL72 needed 8.03 minutes in the same configuration. On Qwen3.5 397B GRPO, Vera Rubin achieved 10.44 minutes on 1,024 GPUs, versus 13.78 minutes for the GB300. With a vLLM backend instead of TensorRT-LLM, the time was 10.94 minutes. For DeepSeek-V3, the company also reports scaling efficiency above 97% when going from 256 to 1,024 GPUs.
On the remaining benchmarks the per-GPU ratio ranges from 1.4x to 1.6x, as shown in the table.
Per-GPU performance of the Vera Rubin NVL72 relative to the GB300 NVL72, MLPerf Training v6.1, Closed division, as given by NVIDIA
| Benchmark | Number of GPUs | Vera Rubin NVL72 vs GB300 NVL72 |
|---|---|---|
| DeepSeek-V3 671B | 1,024 | 2x |
| Qwen3.5 397B GRPO (RL) | 64 | 1.4x |
| gpt-oss 20B | 64 | 1.6x |
| Llama 2 70B LoRA | 72 | 1.5x |
| Llama 3.1 8B | 72 | 1.6x |
| FLUX.1 | 72 | 1.5x |
| DLRMv4 | 128 | 1.4x |
Limits, other participants and the related Endpoints v1.0
MLCommons notes that organizations used the round for different purposes. Chip makers (AMD, Google, NVIDIA) presented new silicon. Server makers showcased different configuration choices and account for 38 systems and 33% of submissions. Cloud providers showed competitive Closed results, with 28% of submissions running on cloud-available systems. Google ran DeepSeek-V3 on 8,192 TPU7x (Ironwood) chips, and Cisco ran it on 64 B300 GPUs, the smallest system to run the benchmark in rounds v6.0 and v6.1. Korea's TTA submitted the round's only energy measurement.
Separately, NVIDIA presented results from MLPerf Endpoints v1.0, which evaluates inference through API endpoints with programming and business workflow workloads. On Kimi K3 (Standardized division, Available category), the Vera Rubin NVL72 with 32 accelerators delivered 41,414.2 tokens/sec in total, with 50.4 tokens/sec per user and power of 100.89 kW. The GB300 NVL72 with 48 accelerators delivered 2,210.4 tokens/sec, with 56.2 tokens/sec per user and 60 kW. From these the company infers more than 10x throughput per kW of provisioned power, with a common minimum of 50 tokens/sec per user. The Vera Rubin results are preliminary, pending verification by MLCommons and may change, while those of the GB300 are verified. On DeepSeek-V4.1-Flash and Qwen3.6-35B-A3B the corresponding per-kW ratios are 9x and 5x.