Baseten
Model deployment and inference · Infrastructure
How they use Nebius AI Cloud
Baseten runs its model-inference stack on Nebius AI Cloud GPU clusters in the United States, Finland and France. Nebius dates this regional deployment to September 2025. Baseten describes a custom text-to-video system running on Nebius capacity for production video generation.
The implementation uses modality-specific kernels, kernel fusion, request prioritization and topology-aware parallelism. Its infrastructure provides intelligent request routing, geographic load balancing, autoscaling and fast cold starts. Baseten also describes continuous batching, asynchronous compute, active-active reliability and multi-cloud capacity management. Baseten reports predictable P90/P99 latency, high utilization and stable uptime on Nebius capacity.
Nebius’s Q2 2025 shareholder letter also states that Baseten runs internal workloads on its infrastructure. The letter describes a completed platform integration that lets end users deploy through Baseten Cloud with Nebius as the underlying infrastructure.
Sources
4 sourcesBaseten: AI video on Nebius
6 Oct 2025 · Baseten · Web page
Mike Bilodeau describes Baseten’s custom text-to-video system on Nebius. Its stack uses continuous batching and asynchronous compute, with active-active reliability and multi-cloud capacity management.
Open original sourceNebius%20-%20Letter%20to%20shareholders%20-%20Q2%202025.pdf?cache-buster=2025-08-26T16:52:55.017Z
14 Oct 2026 · assets.nebius.com · PDF
Nebius states that Baseten runs internal workloads on Nebius. Its Q2 integration also lets end users deploy through Baseten Cloud with Nebius as the underlying infrastructure.
Open original source#partnerships #aiinfrastructure #modelinference | Nebius
14 Oct 2026 · www.linkedin.com · Web page
Nebius states that Baseten Cloud runs on Nebius in the USA, Finland and France. The companies jointly serve model-inference workloads through their integration.
Open original sourceCompany context
baseten.coBaseten builds infrastructure and developer tools for deploying, running, and training AI models. Software and machine learning teams use its platform to serve language, speech, image, and embedding models in production. Its products include Dedicated Inference, ready-to-use Model APIs, managed Training Jobs, and the Loops training SDK, currently in early access. Truss packages custom models, while Chains orchestrates workflows across multiple models. Customers include OpenEvidence, Writer, and Zed. Deployments can run on Baseten’s cloud, in customers’ own clouds, or across both.
Documented financing · private company
Recent updates
31 updatesBaseten launches Reflection Beam inference
Reflection’s Beam became available through Baseten Model APIs and Dedicated Inference. The open-weight model has 501 billion parameters and a one-million-token context window for coding and agent workflows.
BasetenBaseten launches Muse Spark Open inference
Meta’s Muse Spark Open became available through Baseten Model APIs and Dedicated Inference. The 816-billion-parameter model supports text, image, video and audio inputs with a one-million-token context window.
BasetenBaseten adds NVIDIA Nemotron 3.5 Super VL
NVIDIA Nemotron 3.5 Super VL became available through Baseten Model APIs and Dedicated Inference. Developers can also customize the multimodal model through Baseten Training.
Baseten