Nebius and WEKA KV cache benchmark on HGX B300
Company: Nebius
The claim, verbatim
Nebius and WEKA benchmarked a distributed NVMe KV cache tier on NVIDIA HGX B300 in Nebius AI Cloud; WEKA NeuralMesh served 2.4x more requests than HBM-only with a 93% cache hit rate.
Source (primary)
Open-source NVIDIA Run:ai is now available on Nebius AI Cloud Applications - Nebius (-, news_article)
View cached copy (2026-10-16)Live source ↗Archive.org ↗
Quote: “Nebius and WEKA benchmarked a distributed NVMe KV cache tier for agentic inference on NVIDIA HGX B300 in Nebius AI Cloud. Over eight hours on identical hardware, WEKA NeuralMesh with Augmented Memory Grid served 2.4× more requests than HBM-only”
How we checked this
Checked on October 17, 2026. The cited source supports every part of this claim.
The page's related-post summary states all of these figures, though it is a teaser for a separate September 24, 2026 post rather than part of this article's main text.
Confirmed in the source:
- Nebius and WEKA benchmarked a distributed NVMe KV cache tier
- The benchmark ran on NVIDIA HGX B300 in Nebius AI Cloud
- WEKA NeuralMesh served 2.4x more requests than HBM-only
- WEKA NeuralMesh held a 93% cache hit rate
What we did: Read our cached copy of the publisher (https://nebius.com/blog/posts/nvidia-run-ai-open-source-nebius-applications) in full (8,738 characters, retrieved October 16, 2026) and checked each assertion in the claim against it.
Additional evidence
confirms Open-source NVIDIA Run:ai is now available on Nebius AI Cloud Applications - Nebius
Quote: “Nebius and WEKA benchmarked a distributed NVMe KV cache tier for agentic inference on NVIDIA HGX B300 in Nebius AI Cloud. Over eight hours on identical hardware, WEKA NeuralMesh with Augmented Memory Grid served 2.4× more requests than HBM-only”