Official docs state API-Inference is provided free to ModelScope registered users ('API-Inference面向ModelScope注册用户免费提供'), backed by Aliyun compute; usage requires Aliyun account binding with real-name verification, applies dynamic per-model rate limits, and supports 魔粒 deduction tiers (0.5 / 1 / 2 per call by model size).Official docs state that API-Inference is free for ModelScope registered users and backed by Alibaba Cloud compute. Usage requires a real-name-verified Alibaba Cloud binding, dynamic per-model rate limits, and MoLi deduction tiers of 0.5 / 1 / 2 points per call by model size.
关键信息Key details
免费方式Free methodAPI-Inference is free for registered users; dynamic rate limits and 魔粒 (MoLi) deduction applyAPI-Inference is free for registered users; dynamic rate limits and MoLi deduction apply
有效期ValidityOngoing free service; model list iterates as new open models releaseOngoing free service; model list iterates as new open models release
Call open models free via ModelScope API-Inference after binding an Aliyun accountFollow the model-specific steps to register, create a key and connect the model.
✓
注册与接入步骤Registration and access steps
准备条件Prerequisites
ModelScope account
ModelScope account bound to an Aliyun account that has passed real-name verification
Bind your ModelScope account to a real-name verified Aliyun account, then create an access token in ModelScopeBind your ModelScope account to a real-name verified Aliyun account, then create an access token in ModelScope
Send OpenAI-compatible requests to the API-Inference endpoint within the dynamic rate limitsSend OpenAI-compatible requests to the API-Inference endpoint within the dynamic rate limits
每条路径都拆成前置条件、步骤、可复制命令和验证动作；免费条件仍以官方页面实时状态为准。curl 命令按 macOS / Linux shell 编写，Windows 用户可使用同页的 Python 或 Node.js 示例。Each path includes prerequisites, steps, copyable commands and a validation action; free terms still follow the provider's live official policy. The curl examples use a macOS/Linux shell; Windows users can run the Python or Node.js examples on this page.
这些模型出现在平台模型列表里，但用我们的账号真实调用失败。这是账号权限、额度或接入限制，不代表模型能力，也不代表你的账号一定不可用。Listed by the platform but failed when called with our key: an account, quota or access limit, not a model-quality verdict.
deepseek-ai/DeepSeek-V4.1-Flash — 本账号/本端点不可用（无可用服务商），待复核unavailable for our key/endpoint (no serving provider), pending recheck · 记录Record
方法：同一套 22 道固定题、temperature=0、全部自动评分；测试节点：GitHub Actions 托管运行器（美国）。数字只代表该时间点、我们的账号和该测试节点，地区与账号不同结果会不同。Method: the same 22 fixed cases, temperature 0, auto-graded; vantage: GitHub Actions (US). Numbers reflect that run, our accounts and that vantage only.完整报告Full report →
官方来源Official sources
Re-confirmed 2026-09-17 from the server-rendered description of the official limits page: “魔搭通过API-Inference，将开源模型服务化并通过API接口进行标准化，免费提供给广大开发者体验” and “API-Inference 本身为非商业化，非盈利产品”. The detail quotes previously recorded (Aliyun binding + real-name verification, and the 魔粒 tiers 轻量模型 0.5 / 主流模型 1 / 旗舰模型 2 per call) come from the official docs body, which is rendered client-side and could not be re-fetched by our tooling on 2026-09-17 — they still need a manual browser check.Re-confirmed 2026-09-17 from the server-rendered description of the official limits page: “ API-Inference， API ， ” and “API-Inference ， ”. The detail quotes previously recorded (Aliyun binding + real-name verification, and the tiers 0.5 / 1 / 2 per call) come from the official docs body, which is rendered client-side and could not be re-fetched by our tooling on 2026-09-17 — they still need a manual browser check.