@pikku/ai-deepinfra

DeepInfra transcription and speech models for the Vercel AI SDK.

@ai-sdk/deepinfra covers DeepInfra's language, embedding and image models. It does not cover audio, and @ai-sdk/openai-compatible cannot be pointed at DeepInfra's audio endpoints either: they are not OpenAI-shaped. Everything at DeepInfra — chat, ASR, TTS alike — is POST /v1/inference/{model}, and ASR takes its upload under a part named audio rather than OpenAI's file. This package is that gap, and nothing else.

Install

npm install @pikku/ai-deepinfra

Usage

import { createDeepInfra } from '@pikku/ai-deepinfra'
import { VercelAgentRunner } from '@pikku/ai-vercel'

const agentRunner = new VercelAgentRunner({
  deepinfra: createDeepInfra(), // reads DEEPINFRA_API_KEY
})

Then name models the way the runner parses them — provider, slash, model id:

voiceInput({ model: 'deepinfra/openai/whisper-large-v3-turbo' })
voiceOutput({ model: 'deepinfra/hexgrad/Kokoro-82M' })

The runner splits on the first slash only, so the rest reaches DeepInfra intact. The openai/ in that id is the org that published the weights on HuggingFace, not the vendor serving them — this is Whisper running on DeepInfra, with no OpenAI account involved.

For language models, keep @ai-sdk/deepinfra alongside this one; the two are independent.

Options

createDeepInfra({
  apiKey, // defaults to process.env.DEEPINFRA_API_KEY
  baseURL, // defaults to https://api.deepinfra.com/v1/inference
  headers, // merged into every request
  fetch, // injectable, so tests need neither a key nor a network
})

The key is read per request rather than at construction, so a provider can be built before the environment is loaded — the error lands on the call that needed it.

Per-model knobs ride through providerOptions:

await transcribe({
  model: deepinfra.transcription('openai/whisper-large-v3-turbo'),
  audio,
  providerOptions: { deepinfra: { chunk_level: 'word' } },
})

Models

Anything DeepInfra serves; the model id is opaque to this package. Current audio options include:

| Model | Kind | Price | | ------------------------------- | ---- | ------------- | | openai/whisper-large-v3-turbo | ASR | $0.00020/min | | openai/whisper-large-v3 | ASR | $0.00045/min | | Qwen/Qwen3-ASR-0.6B | ASR | $0.00020/min | | Voxtral-Mini-3B | ASR | $0.00100/min | | hexgrad/Kokoro-82M | TTS | $0.80/M chars |

DeepInfra serves streaming-capable ASR models in batch mode only, so their streaming modes are not reachable through this API.

Docs

https://pikku.dev/docs