ModelScope (Alibaba)
partially verifiedpermanent free · inference_host · confidence 70% · mixed · last verified 2026-08-17 · field guide →
Subject record
| Base URLs | https://api-inference.modelscope.cn/v1 openai-compatible |
|---|---|
| API key required | yes |
| Credit card | no |
| Phone verification | unknown |
| May train on prompts | unknown |
| Quota | free_tier [rpd, concurrency] “Official limits page (modelscope.cn, rendered 2026-07-18): "Each registered ModelScope user is currently allowed a total of 2,000 API Inference calls per day, with a maximum of 200 calls per individual model. Specific limits for each model may be dynamically adjusted at any time." Larger models (e.g. deepseek-ai/DeepSeek-R1-0528, deepseek-ai/DeepSeek-V3.2-Exp) are limited to 100 calls/model/day. Concurrency is dynamically rate-limited; quota exposed via modelscope-ratelimit-* response headers (requests-limit 2000, model-requests-limit 200).” |
| Signup friction | steps unknown |
| Env var | MODELSCOPE_API_KEY |
| Notes | Alibaba's community model hub (HF analog) with a free API-Inference tier for registered users. Two official docs fetches (intro, limits) returned only page titles — content is client-side rendered, so all facts are imported from mnfst; method kept as 'imported'. Vision/MLLM models are listed as API-Inference-enabled by mnfst but no specific vision model id could be confirmed. Guide pass 2026-07-17: live limits page (Playwright-rendered) shows 200 calls/model/day (100 for large DeepSeek models) vs the seed-repo 500 ceiling recorded in quota.raw_text — official figure is lower. Response carries modelscope-ratelimit-* headers. Recheck 2026-07-18 (limits page re-rendered via Playwright): 200/model/day and 2,000/day/user CONFIRMED verbatim from official docs; quota.raw_text updated to the official wording and method upgraded imported->doc_fetch. Real-name verification on the linked Alibaba Cloud account also confirmed verbatim ('the corresponding cloud account must have completed real-name verification'). Data policy still unverified — no logging/training statement found on the limits page; CN-jurisdiction flag stands. LIVE PROBE 2026-07-25: key VALID — /models returns 200 with current models (DeepSeek-V3.2/V4-Flash/V4-Pro etc.), but the synthetic completion returned 401 auth_invalid. So the key authenticates for listing yet the free completion path rejects it — likely a modelscope free-tier quirk (per-model activation, or a different completion auth/header). partially_verified is right; the free-completion path needs a manual check before Modelscope is trusted for routing. |
Models (2)
| model | capabilities | context | free | limits |
|---|---|---|---|---|
| Qwen/Qwen3.5-27B | chat | ? | yes | 200 rpd |
| Qwen/Qwen3.5-35B-A3B | chat | ? | yes | 200 rpd |
Provenance
| source_modelscope_docs | official_docs | trust: high | checked 2026-07-18 |
| source_repo_mnfst | curated_repo | trust: medium_high | checked 2026-07-17 |