ModelScope Provider Guide

ModelScope is a Tier-2 catalog provider: OpenAI-compatible, defined by one JSON file (src/lib/providers/catalog/modelscope.json) rather than hand-written code. This page is written from that file.

Verification status: this entry is docs- and roster-verified, not yet live-verified. The fields below come from ModelScope's own public documentation and unauthenticated GET requests — no account was created and no API key was used to build it. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

The docs text on this page was read on 2026-09-29 from the raw document files that the docs site loads, because the rendered pages at modelscope.cn/docs are a JavaScript shell. The English build is https://resouces.modelscope.cn/document/docdata/2026-9-28_15-50-EN/dist/model-service/API-Inference/intro/intro_EN.md (and the matching limits, codewhale, magicube, accounts/registration and accounts/token files); the Chinese build is under https://resouces.modelscope.cn/document/docdata/2026-9-28_15-49-CN/dist/index.json (its model-service/API-Inference/intro/intro_CN.md sibling holds the Chinese intro page). Quotes below are English unless a Chinese build is named.

Key Facts

Provider id: modelscope
Protocol: OpenAI-compatible (/chat/completions) — the intro page (docs) says "ModelScope's API-Inference currently provides OpenAI API-compatible interfaces for large language models."
Base URL: https://api-inference.modelscope.cn/v1
Default model: Qwen/Qwen3.5-35B-A3B
Models in catalog: 12 (selected from the 35-model roster)
Streaming: supported — the intro page's chat example calls client.chat.completions.create(...) with stream=True
Vision: vision: true on Qwen/Qwen3.5-35B-A3B — the intro page's "Vision Models" example calls that id with an image_url content part. It is also true on the other three Qwen3.5 ids (the intro page says "integrated models like Qwen3.5 that support vision") and on Qwen/Qwen3.8-27B, Qwen/Qwen3.8-Flash-Next, MiniMax/MiniMax-M3 and stepfun-ai/Step-3.7-Flash (ModelScope's public model API lists their task as image-text-to-text); see Models
Tool calling: false
Tools while streaming: false
Structured output: false
Structured output + tools together: false — no combined request was sent, because no API key was used
Embeddings: false
Thinking: true — ModelScope's Codewhale integration page (https://modelscope.cn/docs/ecosystem-integrations/api-usage/codewhale) says "Qwen and GLM series models support reasoning_content streaming output"
Billing: free-tier — see Billing below

Quick Start

1. Get an API key

Visit: https://modelscope.cn/home and register or log in with the "Log in/Register" button in the top right corner (https://modelscope.cn/docs/accounts/registration)
Bind an Alibaba Cloud account and complete real-name verification: "After account registration, you must bind your Alibaba Cloud account and complete real-name verification before using API-Inference." (https://modelscope.cn/docs/model-service/API-Inference/intro; binding steps: https://modelscope.cn/docs/accounts/aliyun-binding-and-authorization)
Create an access token at https://modelscope.cn/my/myaccesstoken; the token guide (https://modelscope.cn/docs/accounts/token) has you specify a validity period, "Long-term use" or "Short-term use"
Free for registered users: "API-Inference is provided free of charge to registered ModelScope users." (intro page); the limits page (https://modelscope.cn/docs/model-service/API-Inference/limits) states that API Inference "implements certain limits on usage quotas and concurrency"
Set MODELSCOPE_API_KEY in your .env file

2. Configure

export MODELSCOPE_API_KEY=your-access-token
export MODELSCOPE_MODEL=Qwen/Qwen3.5-35B-A3B   # optional — overrides the default model
export MODELSCOPE_BASE_URL=https://api-inference.modelscope.cn/v1   # optional — proxy or self-hosted gateway

3. Use it

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "modelscope",
  model: "Qwen/Qwen3.5-35B-A3B",
});

console.log(result.content);

# CLI
npx @juspay/neurolink generate "Hello" --provider modelscope

Per-request credentials use the same shape as the other catalog providers:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "modelscope",
  credentials: { modelscope: { apiKey: process.env.MODELSCOPE_API_KEY } },
});

Billing

The entry records free-tier. What the vendor's pages say:

Intro page: "API-Inference is provided free of charge to registered ModelScope users."
Limits page (https://modelscope.cn/docs/model-service/API-Inference/limits): "The free inference API is powered by Alibaba Cloud computing resources." and "Your ModelScope account must first be linked to an Alibaba Cloud account."
Limits page: "Please do not use it for online tasks requiring high concurrency or SLA guarantees". It continues "If you have commercial usage requirements, we recommend using APIs from commercial platforms."
Limits page: "The allowed concurrency for different models will be dynamically rate-limited according to platform load"
Limits page: "API Inference calls can be redeemed by magicubes." with three cost tiers, "Lightweight models: ~0.5 Magicubes per call", "Standard models: ~1 Magicube per call" and "Flagship models: ~2 Magicubes per call".
Magicube page (https://modelscope.cn/docs/magicube/intro): Magicubes can be redeemed for "Civision inference, Civision training, and API Inference services". Its earning table lists "Daily Login Bonus" (action "Sign in", reward "200 Magicubes/day") and "Account Linking Bonus" (action "Link Alibaba Cloud account", reward "50 Magicubes/day").

Models

Model	`vision`	Notes
`Qwen/Qwen3.5-35B-A3B` ⭐	true	NeuroLink default. The id used in the intro page's chat, Anthropic-compatible and Vision Models examples; `models.visionModel`.
`Qwen/Qwen3.5-27B`	true	Fallback. Named in the Chinese intro page's Responses API example and in the LangChain integration page's examples (https://modelscope.cn/docs/ecosystem-integrations/api-usage/langchain). The intro page says "integrated models like Qwen3.5 that support vision". Model API task: `image-text-to-text`.
`deepseek-ai/DeepSeek-V4-Pro-0813`	false	Fallback. Listed on the roster. Model API task: `text-generation`.
`ZhipuAI/GLM-5.2`	false	Fallback. Listed on the roster. Model API task: `text-generation`.
`Qwen/Qwen3.5-397B-A17B`	true	Listed on the roster. The intro page says "integrated models like Qwen3.5 that support vision". Model API task: `image-text-to-text`.
`Qwen/Qwen3.5-122B-A10B`	true	Listed on the roster. The intro page says "integrated models like Qwen3.5 that support vision". Model API task: `image-text-to-text`.
`deepseek-ai/DeepSeek-V4-Flash-0731`	false	Listed on the roster. Model API task: `text-generation`.
`MiniMax/MiniMax-M3`	true	Listed on the roster. Model API task: `image-text-to-text`.
`stepfun-ai/Step-3.7-Flash`	true	Listed on the roster. Model API task: `image-text-to-text`.
`Qwen/Qwen3.8-27B`	true	Listed on the roster. Model API task: `image-text-to-text`.
`mistralai/Mistral-Large-Instruct-2407`	false	Listed on the roster. Model API task: `text-generation`.
`Qwen/Qwen3.8-Flash-Next`	true	Listed on the roster. Model API task: `image-text-to-text`.

The ids come from an unauthenticated GET https://api-inference.modelscope.cn/v1/models call made 2026-09-29 (HTTP 200, 35 models; 12 of them are in this catalog). Each roster entry has the fields id, object (an empty string), owned_by (system) and created. "Model API task" is the Data.Tasks field of an unauthenticated GET https://www.modelscope.cn/api/v1/models/<model id> made 2026-09-29. The catalog does not set a per-model context window, output limit or price. models.defaultContextWindow (32,768) and models.defaultMaxOutputTokens (4,096) are placeholders.

Fallback order when the default is unavailable: Qwen/Qwen3.5-27B → deepseek-ai/DeepSeek-V4-Pro-0813 → ZhipuAI/GLM-5.2. The runtime fallback model name (models.fallbackModelName) is Qwen/Qwen3.5-27B.

Vision: vision: true is set for Qwen/Qwen3.5-35B-A3B, the id the intro page's Vision Models example calls with an image_url content part, for the other Qwen3.5 ids on the basis of the intro page sentence quoted in the table, and for Qwen/Qwen3.8-27B, Qwen/Qwen3.8-Flash-Next, MiniMax/MiniMax-M3 and stepfun-ai/Step-3.7-Flash on the basis of their model API task, image-text-to-text.

Verification status

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for ModelScope — docs- and roster-verified, not live-verified:

Probe	Result
Roster	unauthenticated `GET /v1/models`, HTTP 200, 35 models, 2026-09-29 — no key used
Billing	intro page: "API-Inference is provided free of charge to registered ModelScope users."; limits page: "free of charge", 2026-09-29
Auth-failure shape	unauthenticated `GET /v1/tasks/abc` (the AIGC task-polling route used in the intro page's example) answered HTTP 401 with the JSON body `{"errors":{"message":"Authentication failed, please make sure that a valid ModelScope token is supplied."},"request_id":"..."}`; an unauthenticated `GET /v1/chat/completions` answered HTTP 404 with the plain-text body `404 page not found`, 2026-09-29. `errorRules` holds one rule, status `401` → authentication
Capabilities	`text` and `streaming` come from the intro page's chat example with `stream=True`; `thinking` is `true` on the basis of the Codewhale integration page sentence quoted under Key Facts; `tools`, `structuredOutput` and `embeddings` are `false`; no combined tools+schema request was sent, so `structuredOutputWithTools` is `false`
Live capability sweep	not run. `evidence.liveMatrix` is `null`. Before treating this provider as production-ready, run `npx tsx test/continuous-test-suite-provider-matrix.ts --provider=modelscope` with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting

Each row is a statement from a ModelScope page, quoted from the docs text described at the top of this page.

Situation	What the vendor says	Page
Newly registered account	"After account registration, you must bind your Alibaba Cloud account and complete real-name verification before using API-Inference."	intro
A model id is no longer served	"Over time, as new models are launched, older models may be deprecated and no longer supported. To ensure your API calls work properly, please configure a currently supported model ID."	intro
A model id is no longer served	"As new models are released, older models may gradually be removed from API Inference."	limits
Reasoning models	"Please refer to the API-Inference sample code on the model page as the authoritative source", "especially for reasoning models, as their calling methods may have subtle differences from standard LLMs"	intro
Concurrency or quota limits	"The allowed concurrency for different models will be dynamically rate-limited according to platform load"	limits
Production traffic with concurrency needs	"Please do not use it for online tasks requiring high concurrency or SLA guarantees"	limits

Key Facts​

Quick Start​

1. Get an API key​

2. Configure​

3. Use it​

Billing​

Models​

Verification status​

Troubleshooting​

See also​