Nebius Token Factory provider for the Vercel AI SDK

A community provider for the Vercel AI SDK that gives you access to the open-source models served by Nebius Token Factory (formerly Nebius AI Studio): DeepSeek, Qwen, Llama, GLM, Kimi, gpt-oss and more, plus embedding and image models — all through the standard AI SDK interface.

Installation

npm install nebius-ai-provider ai

Requires Node.js >= 22 and the AI SDK v7.

Setup

Get an API key from the Token Factory console and expose it as NEBIUS_API_KEY:

export NEBIUS_API_KEY="your-api-key"

Quickstart

import { generateText } from 'ai';
import { nebius } from 'nebius-ai-provider';

const { text } = await generateText({
  model: nebius('deepseek-ai/DeepSeek-V4-Flash'),
  prompt: 'Hello!',
});

Model IDs are HuggingFace-style paths. Known models are typed for autocompletion, but any string is accepted, so new models on the platform work immediately. Browse the current catalog at tokenfactory.nebius.com/models/catalog.

Custom provider instance

Use createNebius to customize the API key, base URL, headers, or fetch:

import { createNebius } from 'nebius-ai-provider';

const nebius = createNebius({
  apiKey: process.env.MY_NEBIUS_KEY,
  // baseURL: 'https://api.tokenfactory.nebius.com/v1', // default
  headers: { 'X-Team': 'research' },
});

Streaming

import { streamText } from 'ai';
import { nebius } from 'nebius-ai-provider';

const result = streamText({
  model: nebius('Qwen/Qwen3-235B-A22B-Instruct-2507'),
  prompt: 'Write a haiku about GPUs.',
});

for await (const delta of result.textStream) {
  process.stdout.write(delta);
}

Tool calling

import { generateText, tool, stepCountIs } from 'ai';
import { nebius } from 'nebius-ai-provider';
import { z } from 'zod';

const { text } = await generateText({
  model: nebius('Qwen/Qwen3-32B'),
  tools: {
    weather: tool({
      description: 'Get the weather for a location',
      inputSchema: z.object({ location: z.string() }),
      execute: async ({ location }) => ({ location, temperature: 21 }),
    }),
  },
  stopWhen: stepCountIs(3),
  prompt: 'What is the weather in Amsterdam?',
});

Reasoning models

Nebius serves reasoning models (the DeepSeek-R1 family, Qwen "Thinking" variants) that emit their chain of thought as <think>...</think> blocks. Use the AI SDK's extractReasoningMiddleware to surface reasoning separately from the final text:

import {
  extractReasoningMiddleware,
  generateText,
  wrapLanguageModel,
} from 'ai';
import { nebius } from 'nebius-ai-provider';

const model = wrapLanguageModel({
  model: nebius('deepseek-ai/DeepSeek-R1-0528'),
  middleware: extractReasoningMiddleware({ tagName: 'think' }),
});

const { text, reasoningText } = await generateText({
  model,
  prompt: 'How many primes are there below 100?',
});

console.log(reasoningText); // the model's chain of thought
console.log(text); // the final answer

This works with streamText too — reasoning arrives as reasoning-delta stream parts, separate from text-delta.

Embeddings

import { embed } from 'ai';
import { nebius } from 'nebius-ai-provider';

const { embedding } = await embed({
  model: nebius.embeddingModel('Qwen/Qwen3-Embedding-8B'),
  value: 'sunny day at the beach',
});

Image generation

import { experimental_generateImage as generateImage } from 'ai';
import { nebius } from 'nebius-ai-provider';

const { image } = await generateImage({
  model: nebius.imageModel('black-forest-labs/flux-schnell'),
  prompt: 'A watercolor painting of a lighthouse at dawn',
});

Provider settings

| Option | Type | Default | Description | | --------- | ------------------------ | ---------------------------------------- | -------------------------------------- | | baseURL | string | https://api.tokenfactory.nebius.com/v1 | API base URL | | apiKey | string | process.env.NEBIUS_API_KEY | Bearer API key | | headers | Record<string, string> | — | Extra headers merged into each request | | fetch | typeof fetch | global fetch | Custom fetch for proxies/testing |

Legacy base URLs (AI Studio → Token Factory)

Nebius AI Studio was rebranded to Nebius Token Factory. This provider defaults to the current https://api.tokenfactory.nebius.com/v1 endpoint. If you need one of the legacy endpoints, pass it explicitly:

const nebius = createNebius({
  baseURL: 'https://api.studio.nebius.com/v1', // or https://api.studio.nebius.ai/v1
});

Links

• Nebius Token Factory docs
• Model catalog
• Token Factory cookbook
• Vercel AI SDK docs

License

MIT