The WaveSpeed provider for the Vercel AI SDK gives you access to WaveSpeed's LLMs, image models and video models through the standard AI SDK interfaces (generateText / streamText, generateImage, generateVideo).

Installation

npm install @wavespeed/ai-sdk-provider

Setup

Get an API key from the WaveSpeed dashboard and expose it as the WAVESPEED_API_KEY environment variable:

export WAVESPEED_API_KEY="your-api-key"

Alternatively, pass it explicitly:

import { createWaveSpeed } from '@wavespeed/ai-sdk-provider';

const wavespeed = createWaveSpeed({
  apiKey: 'your-api-key',
});

Language models

Chat models are served by the WaveSpeed LLM gateway (https://llm.wavespeed.ai/v1, override with llmBaseURL) and support streaming, tool calls and structured output. Browse model IDs in the LLM catalog.

import { wavespeed } from '@wavespeed/ai-sdk-provider';
import { generateText, streamText } from 'ai';

const { text } = await generateText({
  model: wavespeed('deepseek/deepseek-v4.1-flash-1017'),
  prompt: 'Write a haiku about GPUs.',
});

const result = streamText({
  model: wavespeed.chat('deepseek/deepseek-v4.1-flash-1017'),
  prompt: 'Explain diffusion models in one paragraph.',
});
for await (const chunk of result.textStream) process.stdout.write(chunk);

Image generation

import { wavespeed } from '@wavespeed/ai-sdk-provider';
import { experimental_generateImage as generateImage } from 'ai';
import { writeFileSync } from 'node:fs';

const { image } = await generateImage({
  model: wavespeed.image('bytedance/seedream-v5.0-pro'),
  prompt: 'A serene mountain lake at sunrise, photorealistic',
  size: '2048x2048',
});

writeFileSync('output.png', image.uint8Array);

The provider submits the prediction, polls WaveSpeed until it completes (1s interval by default, no overall time limit — long-running tasks keep polling until they finish; set pollIntervalMillis / pollTimeoutMillis in createWaveSpeed to change the interval or to opt in to a cap. Each individual HTTP request is bounded (submit 60s, each status poll 30s, output download 120s); abort with abortSignal), and returns the downloaded image bytes.

Video generation

import { wavespeed } from '@wavespeed/ai-sdk-provider';
import { experimental_generateVideo as generateVideo } from 'ai';

const { video } = await generateVideo({
  model: wavespeed.video('wavespeed-ai/minimax-h3/text-to-video'),
  prompt: 'A drone shot flying over a rugged coastline at golden hour',
  aspectRatio: '16:9',
  duration: 5,
});

console.log(video.url);

MiniMax H3 is the cheap open-weights starting point; use wavespeed.video('bytedance/seedance-2.5/text-to-video') for the highest quality.

Video models implement the AI SDK's asynchronous start/status flow, so the AI SDK core orchestrates polling for you.

Provider options

Model-specific inputs that are not part of the standard AI SDK call options can be passed through providerOptions.wavespeed. They are merged into the request body and take precedence over the mapped standard options:

const { image } = await generateImage({
  model: wavespeed.image('bytedance/seedream-v5.0-pro'),
  prompt: 'A cyberpunk city street at night',
  providerOptions: {
    wavespeed: {
      size: '3840*2160',
      enable_sync_mode: false,
    },
  },
});

Consult the model's page on wavespeed.ai for the exact input schema of each model. Fields that a model does not declare in its schema are ignored by the API.

Argument mapping

AI SDK option	WaveSpeed request field
`prompt`	`prompt`
`size` (`{width}x{height}`)	`size` (`{width}*{height}`)
`aspectRatio`	`aspect_ratio`
`seed`	`seed`
`duration` (video)	`duration`
`resolution` (video)	`resolution`
`fps` (video)	`fps`
`generateAudio` (video)	`generate_audio`
image / file inputs	`image` / `images`
`providerOptions.wavespeed.*`	passed through as-is (wins on conflict)