jain.com

DeepSeek-V3 671B trained in 17.65 minutes on 256 Rubin GPUs

Company: CoreWeave

Subject kind
product
Statement date
2026-11-17
Promised amount
17.65 minutes
Current status
stated

The claim, verbatim

CoreWeave trained DeepSeek-V3 671B to target quality in 17.65 minutes on 256 NVIDIA Rubin GPUs across multiple racks, and was the only cloud provider to submit results on this benchmark with NVIDIA Rubin.

Source (primary)

CoreWeave Reports 2.7x Speedup with Vera Rubin Over GB200 on MLPerf DeepSeek-V3 Benchmark - HPCwire (-, news_article)
View cached copy (2026-11-20)Live source ↗Archive.org ↗

Quote: “CoreWeave trained the 671-billion-parameter mixture-of-experts model to target quality in 17.65 minutes on 256 NVIDIA Rubin GPUs”

How we checked this

Checked on November 21, 2026. The cited source supports every part of this claim.

The release gives the time, GPU count and multi-rack setup, and says CoreWeave was the only cloud provider to submit on this benchmark with Rubin.

Confirmed in the source:

What we did: Read our cached copy of the publisher (https://www.hpcwire.com/aiwire/2026/11/17/coreweave-reports-2-7x-speedup-with-vera-rubin-over-gb200-on-mlperf-deepseek-v3-benchmark/) in full (13,089 characters, retrieved November 20, 2026) and checked each assertion in the claim against it.

Additional evidence

confirms CoreWeave Reports 2.7x Speedup with Vera Rubin Over GB200 on MLPerf DeepSeek-V3 Benchmark - HPCwire

Quote: “CoreWeave trained the 671-billion-parameter mixture-of-experts model to target quality in 17.65 minutes on 256 NVIDIA Rubin GPUs”

View cached copy (2026-11-20)Live source ↗Archive.org ↗