Lambda publishes AI training results with a 1.5× software-led speedup
Its language-model submission uses newer NVIDIA hardware. A separate, unofficial image-model run shows the impact of software and tuning; final MLCommons review remains pending.
Loading page…
Its language-model submission uses newer NVIDIA hardware. A separate, unofficial image-model run shows the impact of software and tuning; final MLCommons review remains pending.
Listen to this story
Lambda’s results show how software tuning can shorten training runs without changing the hardware: its post-deadline Flux.1 reproduction improved on a same-hardware benchmark, while its official GPT-OSS-20B submission ran on 64 Blackwell Ultra GPUs. The reported 1.50× Flux.1 gain is not an official MLPerf entry, and Lambda’s submissions remain subject to final MLCommons review, so the comparisons have important status and hardware limits. The results offer teams configurations to investigate, but a faster run does not by itself mean a cheaper one; reserved capacity costs still matter.
Lambda reports a 19.697-minute GPT-OSS-20B run, 2.86× faster than Cisco’s prior H100 result and 1.37× faster than CoreWeave’s B200 result.
On Flux.1, Lambda attributes a 1.27× gain to its software update and a further 1.18× to tuning, including a 1,024 global batch size.
Lambda’s 1.12× comparison with AMD uses its unofficial Flux.1 reproduction against AMD’s official MI355X submission.
Lambda published new AI training results on November 16, showing two routes to shorter runs: newer chips and better software. Its GPT-OSS-20B language-model submission reached the benchmark target in 19.697 minutes. A separate, unofficial Flux.1 image-model run took 26.36 minutes—a claimed 1.5× speedup over the previous round’s fastest result on the same hardware.
The announcement accompanies MLPerf Training v6.1, which the MLCommons results portal dates to November 16, 2026. MLPerf measures elapsed time to train a defined model to a preset quality target, rather than a chip’s theoretical processing speed. Results undergo peer review before publication.
Lambda ran both workloads on 64 NVIDIA Blackwell Ultra GPUs in a single GB300 NVL72 system. Its official GPT-OSS-20B submission used 16 nodes and NVIDIA’s PyTorch Release 26.08. The company says its results remain subject to final MLCommons review and publication; “official” distinguishes the submitted workload from its later Flux.1 reproduction.
For the language-model comparison, Lambda looks back to two other vendors’ submissions in the previous round. It calculates a 2.86× speedup over Cisco’s result using 64 NVIDIA H100 GPUs. Against CoreWeave’s result on 64 Blackwell GPUs across eight HGX B200 systems, Lambda puts the improvement at 1.37×.
GPT-OSS-20B entered the training suite in v6.0 as one of its first mixture-of-experts benchmarks. That design activates only a subset of a model’s parameters for each token, rather than using the entire model. Lambda presents the workload as a reference for teams evaluating this architecture for reasoning and agent applications.
Flux.1 was Lambda’s first multimodal workload run through the MLPerf training harness. Unlike its language-model comparison across hardware generations, this exercise kept the hardware unchanged from the previous round’s benchmark comparison. Lambda attributes the improvement mainly to an updated software stack and changes to the training configuration.
The company describes two successive reproductions. The first updated the software used to run the workload. The second adjusted its learning settings, adding another gain on top. Together, those changes produced the reported 1.50× speedup over the fastest same-hardware result from six months earlier.
Because Lambda completed the reproduction after the submission deadline, it is not an official Flux.1 entry. The company also compares it with AMD’s official result on 64 Instinct MI355X GPUs, claiming a 1.12× speedup. That comparison therefore pairs an unofficial reproduction with an official submission, rather than two entries with the same status.
Lambda’s reproduction write-up links public code and configuration, and says logs are available. Those materials give technical teams a way to examine the settings behind the image-model result. The company recommends checking whether a tested configuration is available for a customer’s own workload before committing to training capacity.
For buyers, Lambda’s own guidance adds a financial qualification: a faster run on more expensive hardware does not automatically mean a lower bill. It recommends comparing elapsed training time with the price and total capacity that must be reserved, then testing a representative workload before making a larger commitment.
Loading discussion...
Join the conversation
Explain when tuning is worth the effort—or when upgrading makes more sense.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.