The full Token Factory Cookbook — 72 recipes across 19 categories, 30 notebooks rendered with their outputs. The runnable backing for the reference architectures on the ecosystem page. Built on open models through one OpenAI-compatible base_url.
Core Token Factory notebooks — adding a LoRA, batch inference, guided JSON, image LoRA.
Point one base_url at Token Factory — the OpenAI-compatible on-ramp, plus aisuite, LiteLLM, and LlamaIndex SDKs.
Production agents across every major framework — LangChain, CrewAI, Agno, AWS Strands, Google ADK, LlamaIndex, Pydantic, and more.
Benchmark Nebius-hosted LLMs on a small data-analysis task and compare their cost, latency, token usage, and correctness.
source ↗See the project readme for benchmark
source ↗This folder has the following agents:
source ↗A production-grade AI Data Analyst Chatbot workflow for n8n, leveraging Nebius Token Factory's DeepSeek-V3-0324-fast LLM for advanced, structured data analysis and Google Workspace automation.
source ↗A simple demonstration of using the Strands library with Nebius Token Factory's API to create an AI assistant that can fetch weather information.
source ↗Short answer: you can get a typed decision with a real probability from an ordinary LLM without generating any text. List the valid answers as letters, ask for one token, and read the probability of each letter from the model's logprobs. On
source ↗A benchmarking tool built with the CAMEL framework that compares the performance of various AI models, including OpenAI and Nebius models. This tool measures and visualizes the speed of different models in terms of tokens processed per seco
source ↗This script runs the same agentic task against multiple Nebius-hosted models and compares token usage, tool calls, wall-clock time, and estimated cost.
source ↗A LangChain Deep Agent that produces decision-grade competitive briefs, powered by an LLM served by Nebius Token Factory, and real-time web research via Tavily.
source ↗A simple yet powerful AI research crew built with CrewAI that leverages multiple specialized agents to discover and analyze groundbreaking technologies.
source ↗Production use case: support teams need agents that combine ticket text, order state, policies, and safe workflow recommendations.
source ↗A SaaS-style customer support agent that answers product questions, looks up real order status, and escalates to a human ticket when it can't confidently resolve the issue.
source ↗Production use case: data teams need read-only agents that can inspect pipelines, query governed datasets, and recommend next actions when metrics break.
source ↗A self-hosted GitHub pull-request reviewer built with LangChain Deep Agents and Nebius Token Factory.
source ↗A powerful AI agent built with Agno that analyzes and provides insights about HackerNews content. This agent uses the Nebius AI model to deliver intelligent analysis of tech news, trends, and discussions.
source ↗Production use case: SRE teams need fast, evidence-backed triage during incidents without giving an LLM the ability to mutate infrastructure directly.
source ↗A multimodal claims-triage agent built on Nebius Token Factory. Drop in a folder of claim documents — in any shape — and the agent reads everything (PDFs, photos, forms), works out what the claim is, confirms its understanding with you, the
source ↗This is a small natural-language-to-SQL data agent inspired by the architecture of eosho/langchaindataagent, but implemented from scratch for this cookbook.
source ↗Two starter examples of a LangChain Deep Agent powered by an LLM served by Nebius Token Factory.
source ↗A starter example of a LangChain Deep Agent powered by an Nvidia Nemotron LLM served by Nebius Token Factory.
source ↗This folder contains production-oriented LangChain examples powered by Nebius Token Factory. Each example is self-contained with its own requirements.txt, env.example, source files, and sample data.
source ↗A powerful AI agent template built with LlamaIndex that demonstrates how to create intelligent agents using the LlamaIndex framework. This starter project implements a Task Management Assistant using the Nebius AI model to showcase LlamaInd
source ↗PitchLoop is a conversational voice-agent example powered by Gradium, Nebius Token Factory, and LangChain.
source ↗A powerful AI agent built with PydanticAI that provides real-time weather information for any city. This agent uses the Nebius AI model to deliver accurate weather forecasts and insights.
source ↗This example shows a sample 'tool calling' agent, built using Google's Agent Development Kit (ADK) framework.
source ↗This agent demonstrates a multi‑step research pipeline powered by Nebius Token Factory models (OpenAI‑compatible) with a FastAPI backend and a Next.js frontend. It uses the model moonshotai/Kimi-K2.6 by default.
source ↗Production use case: security, privacy, and procurement teams need fast vendor reviews grounded in internal controls and contract evidence.
source ↗A Streamlit travel planning app built with LangChain and Nebius. It is inspired by AI travel-agent demos, but it uses a different app structure and planning flow:
source ↗You have a call with a company in 30 minutes. What do you need to know?
source ↗Function calling and live web tools (Tavily) on open models.
Retrieval-augmented generation over PDFs and vector stores — Milvus, Qdrant, Weaviate, LlamaIndex.
A powerful RAG (Retrieval-Augmented Generation) chat application built with Streamlit, LlamaIndex, and Nebius AI's Qwen3 model. This application allows users to upload PDF documents and interact with them through an AI-powered chat interfac
source ↗This example demonstrates how to build a end-to-end RAG pipeline using open source stack.
source ↗Internal Support Agent is a Retrieval-Augmented Generation (RAG) assistant combining web search, document processing, vector retrieval, and intelligent support tools. It features multi‑agent orchestration, automatic support assessment, and
source ↗An AI-powered content generation platform that helps in creating social media posts, articles, and demo applications using a RAG pipeline built with Nebius and Qdrant Cloud vector database. !Qdrant-Nebius RAG
source ↗This example shows querying a PDF using llama index framework and running LLM on Nebius Token Factory
source ↗Train and serve LoRA adapters on open models.
Full fine-tuning pipelines on Token Factory.
Generate training data, distill, and evaluate smaller models.
Model-specific quickstarts — GLM-4.5, GPT-OSS, Qwen3, Nemotron.
Token Factory inside OpenClaw, Tavily, and more.
OpenWiki is a LangChain documentation agent that generates and maintains a Markdown knowledge base from source repositories or connected knowledge sources. This guide configures OpenWiki to use moonshotai/Kimi-K2.7-Code through the Nebius T
source ↗A vision model reading a blurred invoice does not stay silent. It produces a number.
source ↗Run Kimi CLI with Kimi K3 on Nebius Token Factory through its existing OpenAI-compatible provider. This recipe uses Kimi CLI's openailegacy provider, which sends streaming requests to the OpenAI Chat Completions endpoint. It does not add a
source ↗Mattermost Agents can connect to Nebius Token Factory through its built-in OpenAI Compatible service. No Mattermost provider plugin or custom adapter is required.
source ↗OpenClaw can use custom OpenAI-compatible model providers. This recipe registers Nebius Token Factory as a custom provider, keeps the API key outside the configuration file, and shows how to verify the selected model before starting the gat
source ↗OpenHands Agent Canvas is a self-hosted control center for coding agents — a web UI for starting agent conversations, running coding tasks, and setting up automations. Because Nebius Token Factory is a first-class LLM provider in Agent Canv
source ↗pi is a terminal coding agent with a built-in custom-provider configuration path. Nebius Token Factory works through pi's existing openai-completions API mode, so no provider adapter or extension is required.
source ↗Pixeltable is open-source multimodal AI data infrastructure. You declare tables with images, video, audio, documents, and text; computed columns and embedding indexes run incrementally as data arrives. Pixeltable includes a native Nebius To
source ↗Tune an IT access-request agent against environment rewards on Nebius Token Factory, no fine-tuning required.
source ↗Tavily is an LLM-optimized search API built specifically for agents. It delivers clean, structured, and context-aware web results — making it ideal for research, competitive intelligence, and multi-step agentic workflows. Tavily handles que
source ↗Route Token Factory Chat Completions through Respan (formerly Keywords AI) using Respan's existing Custom / Self-hosted provider. The gateway automatically records latency, token, cost, and error traces, so this path needs neither a Token F
source ↗End-to-end applications built on Token Factory.
Coding-agent and code-generation recipes.
Cherry Studio already supports custom OpenAI-compatible providers. This recipe configures its existing OpenAI Chat Completions endpoint for Token Factory. It does not add a provider adapter or enable stateful Responses behavior.
source ↗OpenCode is a coding agent that lives in your terminal, and Nebius Token Factory is one of the providers built into it. There is no gateway to configure and no base URL to paste: run /connect, pick Nebius Token Factory, paste an API key onc
source ↗Use Nebius Token Factory for Zed-owned AI features through Zed's existing OpenAI-compatible provider. No Zed extension, native provider, or Token Factory adapter is required.
source ↗Hands-on workshop material.
A hands-on workshop for building a practical AI agent system with two MCP servers: a Deep Research Agent and a LinkedIn Writing Workflow. The current implementation uses Nebius AI Studio for text LLM calls, Exa for real-time research, Gemin
source ↗---
source ↗Written deep-dives.
Community-contributed guides.
Pelicans on bicycles, snake games — open models having fun.
Point your webcam at something and get a running commentary from a vision model. It is exactly as fun as it sounds, and it takes about thirty seconds to get going.
source ↗This is inpired by Simon Willison's experiment called pelican riding a bicycle. It has become a cool vibe benchmark to evaluate LLMs. Original github repo: simonw/pelican-bicycle
source ↗A visual snake battle game where two LLMs compete against each other! Watch as AI models control their snakes in a battle of survival.
source ↗No recipes match.