@runtypelabs/mixlayer-ai-provider weekly downloads statistics

@runtypelabs/mixlayer-ai-provider

An AI SDK provider for Mixlayer, which serves open-weight models over an OpenAI-compatible endpoint. Mixlayer's current catalog includes Qwen and GLM models. This provider accepts any model id and adds family-specific behavior only for models it recognizes.

It wraps @ai-sdk/openai-compatible for Chat Completions and @ai-sdk/openai for Responses API models, baking in everything Mixlayer needs to behave correctly out of the box:

• the Mixlayer base URL
• the Qwen thinking toggle (Chat Completions thinking field / Responses reasoning.effort), scoped to the Qwen 3.5 / 3.6 models
• reasoning support — <think> tags and native reasoning_content both surface as AI SDK reasoning parts
• image input through standard AI SDK file parts on vision-capable models
• Responses API models and a Responses WebSocket fetch adapter for lower latency streaming on repeat calls
• tolerant model-id handling (strips a leading mixlayer/ prefix)

Install

pnpm add @runtypelabs/mixlayer-ai-provider ai@^7.0.127

This package requires AI SDK 7.0.127 or later in the v7 series and is ESM-only. For Node.js applications, use Node.js 22+. The package avoids Node-only WebSocket dependencies so it can be bundled for standards-based runtimes such as Cloudflare Workers. ai is a peer dependency, so your app dedupes a single AI SDK version.

Usage

import { mixlayer } from '@runtypelabs/mixlayer-ai-provider'
import { streamText } from 'ai'

const result = streamText({
  model: mixlayer('qwen/qwen3.8-27b'),
  prompt: 'Explain reasoning models in one paragraph.',
})

for await (const text of result.textStream) process.stdout.write(text)

The default mixlayer instance reads MIXLAYER_API_KEY from the environment (Node). To set the key (or any other option) explicitly, create your own provider with createMixlayer:

import { createMixlayer } from '@runtypelabs/mixlayer-ai-provider'

const provider = createMixlayer({
  apiKey: process.env.MIXLAYER_API_KEY,
  thinking: false, // disable Qwen 3.5 / 3.6 thinking
})

const model = provider('qwen/qwen3.5-9b')

Vision

Models in MIXLAYER_VISION_MODEL_IDS accept image inputs through standard AI SDK file parts. The same message shape works with Chat Completions and Responses models:

import { readFile } from 'node:fs/promises'
import { generateText } from 'ai'
import { mixlayer } from '@runtypelabs/mixlayer-ai-provider'

const image = await readFile('./diagram.png')
const result = await generateText({
  model: mixlayer('qwen/qwen3.6-27b'),
  messages: [
    {
      role: 'user',
      content: [
        { type: 'text', text: 'What does this diagram show?' },
        { type: 'file', mediaType: 'image/png', data: image },
      ],
    },
  ],
})

console.log(result.text)

Mixlayer also accepts supported remote image URLs. See the Mixlayer vision guide for current formats and limits. Text-only models return a media error when sent an image.

Responses API and WebSocket mode

The callable provider and provider.languageModel(id) use Chat Completions by default for backwards compatibility. Use provider.responses(id) for Mixlayer's OpenAI-compatible Responses API:

import { createMixlayer } from '@runtypelabs/mixlayer-ai-provider'
import { streamText } from 'ai'

const provider = createMixlayer({
  apiKey: process.env.MIXLAYER_API_KEY,
  thinking: false, // for Qwen 3.5 / 3.6, maps to reasoning.effort: 'none'
})

const result = streamText({
  model: provider.responses('qwen/qwen3.5-9b'),
  prompt: 'Say hello via the Responses API.',
})

To route streaming Responses API requests over Mixlayer's WebSocket endpoint, pass createMixlayerWebSocketFetch() as the provider fetch:

import {
  createMixlayer,
  createMixlayerWebSocketFetch,
} from '@runtypelabs/mixlayer-ai-provider'
import { streamText } from 'ai'

const wsFetch = createMixlayerWebSocketFetch()
const provider = createMixlayer({
  apiKey: process.env.MIXLAYER_API_KEY,
  thinking: false,
  fetch: wsFetch,
})

try {
  const result = streamText({
    model: provider.responses('qwen/qwen3.5-9b'),
    prompt: 'Say hello over WebSocket.',
  })

  for await (const text of result.textStream) process.stdout.write(text)
} finally {
  wsFetch.close()
}

createMixlayerWebSocketFetch() is designed to be Workers-friendly by default: it performs a fetch() request with Upgrade: websocket, uses Response.webSocket, and calls accept() before sending the response.create event. That matches Cloudflare Workers' outbound WebSocket API and keeps this package free of a hard ws / Node.js dependency.

Runtimes that do not expose Response.webSocket on upgraded fetch responses (for example, typical Node.js fetch implementations) can pass a custom connect option. Browsers also cannot attach arbitrary Authorization headers to direct WebSocket handshakes, so browser apps should use a server/Worker connector or proxy rather than connecting directly with a secret API key.

If you want provider registries or provider(id) to use Responses API models, set defaultModelApi: 'responses' in createMixlayer(...).

The provider also slots into createProviderRegistry:

import { createProviderRegistry } from 'ai'

const registry = createProviderRegistry({ mixlayer })
const model = registry.languageModel('mixlayer:qwen/qwen3.6-27b')

Models…