DeepInfra — ai-lcr Docs

DeepInfra — open-weights host. The one quirk that trips everyone — the chat endpoint lives at /v1/openai, so baseURL must end in /v1/openai.

DeepInfra hosts open-weights models (DeepSeek, MiniMax, Kimi, …) cheaply.

The one quirk that trips everyone: DeepInfra serves the OpenAI-compatible API at /v1/openai/chat/completions — the /v1/ comes before openai. Since createOpenAICompatible appends /chat/completions to your baseURL, set the base to …/v1/openai:

import { createOpenAICompatible } from "@ai-sdk/openai-compatible";

const deepinfra = createOpenAICompatible({
  name: "deepinfra",
  baseURL: "https://api.deepinfra.com/v1/openai", // ← not /v1
  apiKey: process.env.DEEPINFRA_API_KEY,
});
// HuggingFace-style ids: deepinfra("deepseek-ai/DeepSeek-V4-Flash")

Model ids are HuggingFace-style paths (deepseek-ai/DeepSeek-V4-Flash). DeepInfra returns usage.estimated_cost on each response — handy for reconciling what you were actually billed against the cost you tracked.