The vLLM project has used Nebius GPU clusters to test and optimize inference, including work on DeepSeek R1. Nebius says the project used its non-replicated disks to store model weights. Nebius says these disks allowed vLLM to load large model weights quickly without bottlenecks during initialization. A reproducible inference benchmark repository reports that the vLLM team runs benchmarks on Nebius H200 machines. Nebius also reports high-throughput inference, seamless scalability and integration of multi-latent attention and multi-token prediction. Its customer story describes RLHF workloads for synthetic data generation, reward modeling and real-time inference evaluation on Nebius compute clusters.
vLLM also acknowledges Nebius as a compute-credit sponsor in its account of CI and benchmarking. An April 2026 post by Narek T. describes Nebius helping vLLM adopt Blackwell GPUs as part of its community support. Its roadmap includes broader support for very large models and distributed inference. Nebius's customer story states that vLLM plans to continue using its infrastructure for inference development.
Nebius reports zero hardware-related issues and consistently accurate hardware performance metrics. TD SYNNEX describes 24/7 access to Nebius hardware.
Nebius vLLM customer story28 Sept 2026 · Nebius · Web pageThe customer story describes vLLM compute-cluster experiments, inference benchmarks, DeepSeek R1 work and storage.
Nebius Token Factory | Open-Source AI Models Platform
18 Oct 2026 · www.tdsynnex.com · Web page
TD SYNNEX says vLLM used Nebius compute clusters to test, benchmark and optimize inference, including DeepSeek R1. It reports zero hardware-related issues, consistently accurate hardware performance metrics and 24/7 hardware access.
Nebius Token Factory | Open-Source AI Models Platform18 Oct 2026 · www.tdsynnex.com · Web pageTD SYNNEX says vLLM used Nebius compute clusters to test, benchmark and optimize inference, including DeepSeek R1. It reports zero hardware-related issues, consistently accurate hardware performance metrics and 24/7 hardware access.
GitHub - Michaelvll/llm-ie-benchmarks: A collection of reproducible inference engine benchmarks · GitHub
3 Oct 2026 · github.com · Web page
Michaelvll's benchmark repository says the vLLM team runs inference benchmarks on Nebius H200 machines. It provides SkyPilot commands targeting Nebius and DeepSeek-R1 benchmark results.
GitHub - Michaelvll/llm-ie-benchmarks: A collection of reproducible inference engine benchmarks · GitHub3 Oct 2026 · github.com · Web pageMichaelvll's benchmark repository says the vLLM team runs inference benchmarks on Nebius H200 machines. It provides SkyPilot commands targeting Nebius and DeepSeek-R1 benchmark results.
Nebius April 2025 digest
7 May 2025 · Nebius · Web page
Nebius calls vLLM its customer and describes the project testing, benchmarking and optimizing inference using Nebius infrastructure.
Nebius April 2025 digest7 May 2025 · Nebius · Web pageNebius calls vLLM its customer and describes the project testing, benchmarking and optimizing inference using Nebius infrastructure.
vLLM · compute-credit acknowledgment
16 Jul 2026 · vLLM project · Web page
The vLLM project thanks Nebius among the partners sponsoring it with compute credits in its article about CI and release testing.
vLLM · compute-credit acknowledgment16 Jul 2026 · vLLM project · Web pageThe vLLM project thanks Nebius among the partners sponsoring it with compute credits in its article about CI and release testing.
vLLM is a community-developed open-source engine for running and serving large language models. Developers, researchers and infrastructure teams use its Python library, offline batch inference tools and OpenAI-compatible API server to deploy models across GPUs, CPUs and other accelerators. Its features include PagedAttention memory management, continuous batching, prefix caching, quantization and distributed inference. Originally developed at UC Berkeley, the project is now hosted by the PyTorch Foundation and supports both research workloads and production applications.
vLLM released support for Reflection’s 501-billion-parameter Beam model, including reasoning, tool calling and structured outputs. A dedicated Docker image provides deployment on NVIDIA Blackwell GPUs.
vLLM adds launch-day support for Nemotron 3.5 Super VL
vLLM announced support for NVIDIA's new multimodal model, with BF16 and NVFP4 deployment options. The model accepts text, images, documents and video for agent workflows.
vLLM releases launch-day support for Muse Spark Open
vLLM worked with Meta Superintelligence Labs to support Muse Spark Open, an open-weight model for agentic coding. Prebuilt NVIDIA and AMD images were available, with nightly wheel inclusion pending review and merging.
The vLLM project has used Nebius GPU clusters to test and optimize inference, including work on DeepSeek R1. Nebius says the project used its non-replicated disks to store model weights. Nebius says these disks allowed vLLM to load large model weights quickly without bottlenecks during initialization. A reproducible inference benchmark repository reports that the vLLM team runs benchmarks on Nebius H200 machines. Nebius also reports high-throughput inference, seamless scalability and integration of multi-latent attention and multi-token prediction. Its customer story describes RLHF workloads for synthetic data generation, reward modeling and real-time inference evaluation on Nebius compute clusters.
vLLM also acknowledges Nebius as a compute-credit sponsor in its account of CI and benchmarking. An April 2026 post by Narek T. describes Nebius helping vLLM adopt Blackwell GPUs as part of its community support. Its roadmap includes broader support for very large models and distributed inference. Nebius's customer story states that vLLM plans to continue using its infrastructure for inference development.
Nebius reports zero hardware-related issues and consistently accurate hardware performance metrics. TD SYNNEX describes 24/7 access to Nebius hardware.
Nebius vLLM customer story28 Sept 2026 · Nebius · Web pageThe customer story describes vLLM compute-cluster experiments, inference benchmarks, DeepSeek R1 work and storage.
Nebius Token Factory | Open-Source AI Models Platform
18 Oct 2026 · www.tdsynnex.com · Web page
TD SYNNEX says vLLM used Nebius compute clusters to test, benchmark and optimize inference, including DeepSeek R1. It reports zero hardware-related issues, consistently accurate hardware performance metrics and 24/7 hardware access.
Nebius Token Factory | Open-Source AI Models Platform18 Oct 2026 · www.tdsynnex.com · Web pageTD SYNNEX says vLLM used Nebius compute clusters to test, benchmark and optimize inference, including DeepSeek R1. It reports zero hardware-related issues, consistently accurate hardware performance metrics and 24/7 hardware access.
GitHub - Michaelvll/llm-ie-benchmarks: A collection of reproducible inference engine benchmarks · GitHub
3 Oct 2026 · github.com · Web page
Michaelvll's benchmark repository says the vLLM team runs inference benchmarks on Nebius H200 machines. It provides SkyPilot commands targeting Nebius and DeepSeek-R1 benchmark results.
GitHub - Michaelvll/llm-ie-benchmarks: A collection of reproducible inference engine benchmarks · GitHub3 Oct 2026 · github.com · Web pageMichaelvll's benchmark repository says the vLLM team runs inference benchmarks on Nebius H200 machines. It provides SkyPilot commands targeting Nebius and DeepSeek-R1 benchmark results.
Nebius April 2025 digest
7 May 2025 · Nebius · Web page
Nebius calls vLLM its customer and describes the project testing, benchmarking and optimizing inference using Nebius infrastructure.
Nebius April 2025 digest7 May 2025 · Nebius · Web pageNebius calls vLLM its customer and describes the project testing, benchmarking and optimizing inference using Nebius infrastructure.
vLLM · compute-credit acknowledgment
16 Jul 2026 · vLLM project · Web page
The vLLM project thanks Nebius among the partners sponsoring it with compute credits in its article about CI and release testing.
vLLM · compute-credit acknowledgment16 Jul 2026 · vLLM project · Web pageThe vLLM project thanks Nebius among the partners sponsoring it with compute credits in its article about CI and release testing.
vLLM is a community-developed open-source engine for running and serving large language models. Developers, researchers and infrastructure teams use its Python library, offline batch inference tools and OpenAI-compatible API server to deploy models across GPUs, CPUs and other accelerators. Its features include PagedAttention memory management, continuous batching, prefix caching, quantization and distributed inference. Originally developed at UC Berkeley, the project is now hosted by the PyTorch Foundation and supports both research workloads and production applications.
vLLM released support for Reflection’s 501-billion-parameter Beam model, including reasoning, tool calling and structured outputs. A dedicated Docker image provides deployment on NVIDIA Blackwell GPUs.
vLLM adds launch-day support for Nemotron 3.5 Super VL
vLLM announced support for NVIDIA's new multimodal model, with BF16 and NVFP4 deployment options. The model accepts text, images, documents and video for agent workflows.
vLLM releases launch-day support for Muse Spark Open
vLLM worked with Meta Superintelligence Labs to support Muse Spark Open, an open-weight model for agentic coding. Prebuilt NVIDIA and AMD images were available, with nightly wheel inclusion pending review and merging.
