Vercel AI SDK
The Vercel AI SDK reaches MartLLM through its official
@ai-sdk/openai-compatible provider. Every chat model in the
catalog works with generateText and
streamText, and image models work with generateImage.
The examples below were run against the live API with ai 7.0 and
@ai-sdk/openai-compatible 3.0.
Install
npm i ai @ai-sdk/openai-compatible zod
Create an API key in Dashboard → Keys and
put it in UBERLLM_API_KEY. Accounts start at $0, so add credit before the
first call.
Set up the provider
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
export const uberllm = createOpenAICompatible({
name: "uberllm",
apiKey: process.env.UBERLLM_API_KEY,
baseURL: "https://api.martllm.com/v1",
includeUsage: true, // token usage on streamed responses too
supportsStructuredOutputs: true, // needed for typed output (Output.object): sends your JSON schema to the model
});
A complete Next.js chat app with this provider, a tool and typed output is in the Starter kits.
apiKey is sent as Authorization: Bearer <key>. Model ids are the ids on
martllm.com/models, for example
deepseek/deepseek-v4-flash, anthropic/claude-sonnet-5.5 or
openai/gpt-5.5.
Generate text
import { generateText } from "ai";
const { text, usage } = await generateText({
model: uberllm("deepseek/deepseek-v4-flash"),
prompt: "Say hello in one short sentence.",
});
console.log(text, usage);
usage.raw.cost_usd is what the call cost, in US dollars.
Stream
import { streamText } from "ai";
const result = streamText({
model: uberllm("deepseek/deepseek-v4-flash"),
prompt: "Count from 1 to 5.",
});
for await (const delta of result.textStream) {
process.stdout.write(delta);
}
console.log(await result.usage);
Tool calls
import { generateText, tool, stepCountIs } from "ai";
import { z } from "zod";
const { text, steps } = await generateText({
model: uberllm("deepseek/deepseek-v4-flash"),
tools: {
weather: tool({
description: "Get the current temperature in a city, in Celsius",
inputSchema: z.object({ city: z.string() }),
execute: async ({ city }) => ({ city, celsius: 21 }),
}),
},
stopWhen: stepCountIs(3), // call the tool, then answer
prompt: "What is the temperature in Paris right now?",
});
console.log(text);
Most chat models take tools. A model page on martllm.com says when one does not.
Images
Image models use uberllm.imageModel(...) with generateImage. The SDK asks
for b64_json, so the picture comes back in the answer.
import { generateImage } from "ai";
const { image } = await generateImage({
model: uberllm.imageModel("black-forest-labs/flux.2-klein-4b"),
prompt: "A red bicycle on a white background",
size: "1024x1024",
});
// image.uint8Array holds the PNG
Image ids are listed under Images on martllm.com/models.
Editing models (ids ending in -edit) take POST /v1/images/edits, which
@ai-sdk/openai-compatible does not call. Use the REST call in
Make a video or image for those.
Video and music
@ai-sdk/openai-compatible has no video or music model. Video and music are jobs:
start one with POST /v1/videos (or /v1/music), then poll it or get a
callback. See Make a video or image, or let an agent do it through
the MCP server.
Prices
Each call costs what its provider charged for it. No MartLLM fee on requests.
Adding credit by card costs 5.5% + $0.50. To cap what a chat call can cost, pass
provider.max_price in the request body through providerOptions:
await generateText({
model: uberllm("deepseek/deepseek-v4-flash"),
prompt: "Hi",
providerOptions: { uberllm: { provider: { max_price: { prompt: 0.2, completion: 0.4 } } } },
});
See Pricing and Models & routing.