OpenClaw LaunchOpenClaw Launch
← Home

Guide

OpenClaw + ModelScope: Free Inference, With Two Catches

ModelScope is Alibaba's model hub, and the part people are searching for is its inference API — OpenAI-compatible, free at meaningful volume, and stocked with Qwen3.5, DeepSeek-V4, GLM and MiniMax. It works with an OpenClaw agent through the custom provider path. Here is the setup, and the two things worth knowing before you plan around it.

What ModelScope Gives You

PropertyValue
Base URLhttps://api-inference.modelscope.cn/v1
FormatOpenAI-compatible chat completions
Free quota2,000 requests/day total, up to 500/day per model
ModelsQwen3.5 family, DeepSeek-V4, GLM-4.7-Flash, GLM-5.2, MiniMax, Mistral
Context8K to 1M tokens, depending on the model
Card requiredNo

Two thousand requests a day at no cost is genuinely generous — comfortably more than a personal assistant bot gets through — and the model list is current rather than a museum of last year's releases.

Catch One: Registration Needs a Chinese Phone or Alibaba Account

This is the wall most people hit. Signing up at modelscope.cn requires an Alibaba account or a Chinese phone number. If you have neither, this route is closed to you regardless of how good the free tier is, and no amount of configuration will change that. Read this page as “can I use it” before “how do I use it.”

If you cannot register, the equivalent free-ish routes are OpenRouter (which carries some free model slots) or running weights locally with Ollama. The same Qwen models are available both ways — see OpenClaw + Qwen.

Catch Two: The Endpoint Is in China

The API is served from mainland China, and our agent containers run in Germany. Every turn crosses that distance twice, which shows up as latency on top of whatever the model itself takes. It is entirely usable for a chat assistant and noticeably slow for a long agentic loop making many tool calls in sequence.

Consider where your data is going as well. Prompts sent to this endpoint are processed under Chinese jurisdiction. For a personal bot that may be irrelevant; for anything carrying colleagues' or customers' messages it is a decision to make deliberately rather than by accident.

Setup on OpenClaw Launch

  1. Sign in at modelscope.cn, open Access Tokens, and create a new token.
  2. On your API Keys page, scroll to the Custom API Provider card.
  3. Set API Endpoint URL to https://api-inference.modelscope.cn/v1 — base URL only, no /chat/completions suffix. Including it causes a 404.
  4. Paste the token into API Key and save.
  5. Set the model to the exact ID ModelScope lists on the model's page, such as Qwen/Qwen3.5-32B-Instruct. The publisher prefix is part of the ID.

The full custom-provider walkthrough, including what to do when a provider names its fields differently, is in the custom endpoint guide.

Self-Hosted Configuration

Running your own OpenClaw, ModelScope is an OpenAI-compatible provider block:

{
  "models": {
    "providers": {
      "modelscope": {
        "baseURL": "https://api-inference.modelscope.cn/v1",
        "apiKey": "ms-..."
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "modelscope/Qwen/Qwen3.5-32B-Instruct"
      }
    }
  }
}

Is the Free Tier Enough for an Always-On Bot?

For one person, usually yes. For a bot in a busy group chat, watch the per-model cap: 500 requests per day against a single model runs out faster than the 2,000 total suggests, and agentic turns spend several requests each rather than one. If you are running scheduled jobs on top of chat, budget accordingly — see scheduled agent tasks.

What running out looks like. The endpoint starts returning errors rather than degrading politely, and to a chat user that reads as “the bot is broken.” If your agent needs to be dependable, a free quota is a nice supplement to a paid provider rather than a replacement for one.

FAQ

Is the ModelScope API really free?

Yes, within quota — 2,000 requests a day in total and up to 500 per individual model, with no card required. The cost is the registration requirement and the latency, not money.

Can I use ModelScope without a Chinese phone number?

Only if you already have an Alibaba account. That is the documented registration path and there is no international bypass. Without one, use OpenRouter or run the same open models locally with Ollama.

What is the ModelScope API base URL?

https://api-inference.modelscope.cn/v1. Enter only that base — OpenClaw appends /chat/completions itself.

Which models are worth using on it?

For agent work, prefer models with a solid tool-calling record — the Qwen3.5 instruct family and GLM-5.2 are the reasonable picks. A model that answers well in chat but ignores tool schemas will fail quietly as an agent.

What's Next?