Serverless Cookbook

Runnable, workload-first examples for Serverless AI Jobs and Endpoints on Nebius.

This repository shows how to run real AI/ML workloads without managing VM lifecycle directly.
Examples focus on practical use cases such as:

• model training
• fine-tuning
• batch inference
• LLM serving
• simulations and domain workloads

⚠️ This is a community-style repository maintained by Nebius engineers.
It is not official product documentation, but reflects real usage patterns and experiments.
APIs and behavior may evolve.

For official documentation, see: https://docs.nebius.com/serverless

Try a workload

Pick the section that matches your goal — each links to runnable examples:

• 🚀 Quickstarts — lowest-friction first runs.
• 🏋️ Training — model training and fine-tuning workloads.
• ⚡ Inference — endpoint serving and batch inference workloads.
• 🔁 MLOps / Pipelines - orchestration, artifact handoffs, and multi-stage workflows.
• 🧬 Life Science — domain-specific simulation and analysis workloads.
• 🤖 Robotics — simulation, dataset generation, and robotics workflows.

Getting started

1. Install the Nebius CLI: Install guide
2. Configure your CLI profile and project: Configure guide
3. Pick an example from the sections below
4. Follow the example README and verify the expected output
   • Optional: shared setup helpers live in scripts/README.md

Example catalog

🚀 Quickstarts

Lowest-friction first runs.

• first-job.md — run nvidia-smi in a Serverless AI job
• first-endpoint.md — deploy a quick nginx endpoint

🏋️ Training

Model training and fine-tuning workloads.

• axolotl-finetuning — get started fine-tuning with Axolotl
• train-and-serve — fine-tune TinyLlama in a Job and serve it with a vLLM Endpoint

⚡ Inference

Endpoint serving and batch inference workloads.

• vllm-endpoint — serve Qwen with an OpenAI-compatible vLLM endpoint

🔁 MLOps / Pipelines

Workflow orchestration and artifact handoff patterns.

• video-transcription-pipeline - orchestrate Object Storage, CPU jobs, and GPU Whisper jobs with Prefect

🧬 Life Science

Domain-specific simulation and analysis workloads.

• openmm-simulation — run GPU-backed molecular dynamics simulations with OpenMM

🤖 Robotics

Robotics and physical-AI experiment loops.

• lerobot-finetune-job — fine-tune a LeRobot ACT or Diffusion policy on a robotics dataset in a serverless GPU job
• smolva-ft-norma-core — fine-tune SmolVLA for SO-101 with bundled trajectories

Awesome Community Projects

External examples and writeups from the community running serverless workloads on Nebius. Got something to add? Open a PR.

Robotics

• 🤖 Positronic + Nebius serverless workflows — Convert datasets, train ACT/SmolVLA, and serve checkpoints as endpoints — all serverless on Nebius. — by vertix · 💻 code
• 🦾 norma-core SmolVLA — Nebius fine-tune recipe — Upstream recipe the robotics/smolva-ft-norma-core example mirrors. — by norma-core · 💻 code

Repository structure

serverless-cookbook/
├─ README.md
├─ CONTRIBUTING.md
├─ DEVELOPER_GUIDE.md
├─ LICENSE
├─ quickstarts/
├─ training/
├─ robotics/
├─ inference/
├─ mlops/
├─ life-science/
├─ robotics/

Resources

• Contributing
• Developer Guide
• Serverless AI overview docs
• CLI AI reference