ModelScope

modelscope.cn · Compute · 3 records · JSON

ModelScope Computetrialnational-idno cardphoneRretention-clamped
ModelScope Notebook Computefails: identitynational-idno cardRretention-clamped
ModelScope Studio Computefails: identitycountry-phoneno cardphoneU

ModelScope Compute

trialnational-idno cardphoneRretention-clamped

modelscope.cn/docs/notebook/intro · GPU cloud · compute/modelscope · JSON · compare

Free tier	" Notebook 36 " - 36 GPU hours as a new-user initial grant, while the CPU environment (8 cores, 32 GB) stays free with no stated limit. unit hours; metered on wall-clock; one-time-credit
Price	none stated
Per-object cap	not applicable
Retention	the 36-hour grant does not refill and a single instance may run at most 10 hours, so a 6-hour idle timeout spends the grant without producing anything
Egress	not stated
Identity	national-id a bound Alibaba Cloud account is required, and that account needs real-name verification: one source states "binding an Alibaba Cloud account is required and that account requires real-name verification", and the grant is "36 per new user after binding an Alibaba Cloud account". The record does not state which document that verification accepts — a Chinese national ID or a passport — whether a card or phone is also needed on either account, any hold, or debit or prepaid acceptance
Limits	retention-clamped the 36-hour grant does not refill and a single instance may run at most 10 hours, so a 6-hour idle timeout spends the grant without producing anything source the 36-hour grant does not refill and a single instance may run at most 10 hours, so a 6-hour idle timeout spends the grant without producing anything
Binding	retention-clamped the 36-hour grant does not refill and a single instance may run at most 10 hours, so a 6-hour idle timeout spends the grant without producing anything
Interface	rest inferred from the tooling the provider documents; not probed
Tooling	not stated
Composes with	The GPU grant is one-time, but the perpetual 8-core CPU notebook with 100 GB of storage is a free place to hold data and run scraping or embedding jobs, which is the part that composes with a storage provider.
Reliability	no reliability event recorded
Notes	The official page could not be read directly: the reader returned 'no content' and the Wayback replay of the 2026-08-03 capture was empty, so this row is `reported` on three independent quotations rather than `documented`. The headline number in circulation has changed and the direction is downward: guides say 100 hours, the doc says 36, and one author explicitly corrects the older figure. Two adjacent lanes are separate products and are not added to this row: free xGPU on (shared GPU for Spaces, requires joining an xGPU-Explorers organisation, approval-gated, with per-user duration caps), and the AMD GPU incentive programme (up to 1000 hours, requires an AMD developer account plus publishing activity).
Evidence rung	reported
Checked	2026-10-22 (live); stale on 2026-11-05
Parsed numbers	`$.limits[0].value` limit:retention-clamped 10.0 hours

Evidence (4)

date	source	what it proved
2026-10-22	www.3520.net/free-cargo/18257.html	quoted from the official Notebook doc: 36 per new user after binding an Alibaba Cloud account; CPU 8 cores / 32 GB free long term; 100 GB storage; single instance at most 10 hours; auto-stop after 1 idle hour; the author corrects the widely repeated '100 hours at signup' figure as stale
2026-10-22	lanmao.ai/zh/o/modelscope-notebook-free-gpu/	quoted, checked 2026-10-02: 36 hours is a one-time new-user GPU grant; binding an Alibaba Cloud account is required and that account requires real-name verification
2026-10-22	tianchi.aliyun.com/forum/post/570847	Alibaba's own Tianchi forum post states the free instance is 36 hours of single-card A10 compute
2026-10-22	modelscope.cn/docs/notebook/intro	official doc page returned 422 no-content from the reader and the 2026-08-03 Wayback capture replayed empty; the figure rests on three sources quoting the doc, not on the doc itself

More fields (1)

{
 "wave2": {
  "class": "recurring-notebook-quota",
  "free_tier_kind": "one-time-credit",
  "shape": {
   "vcpu": 8,
   "ram_gib": 32,
   "gpu_class": "A10 24GB",
   "gpu_count": 1
  }
 }
}

Unknown or withheld values (7)

$.egress.metered egress-unknown
$.egress.terms not-stated
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed
$.retention.notes process
$.tooling not-stated

ModelScope Notebook Compute

fails: identitynational-idno cardRretention-clamped

www.modelscope.cn/learn/433641 · notebook runtime · compute/modelscope-notebook · JSON · compare

Free tier	ModelScope and AMD run a developer incentive programme. Step 0 is registering for the AMD AI Developer Program through the ModelScope Notebook page, which pays 100 hours of AMD GPU notebook time automatically. Content paths stack on top: 25 hours per approved technical article capped at 50, 50 hours per approved notebook capped at 150, and 700 hours or more for an approved deployed application, which the vendor states totals 1,000 hours or more. The AMD shape reported is 8 cores, 200 GB RAM and 192 GB of VRAM, which is MI300X class. unit hours; metered on wall-clock; one-time-credit
Price	none stated
Per-object cap	not applicable
Retention	the 100 hours are a one-time grant that only ticks while the instance runs, a single boot is capped at about 8 to 10 hours, and only files under /workspace persist, so the 192 GB card is reachable in 8-hour slices; the harder wall is that the ModelScope notebook needs a real-name-verified Alibaba Cloud account linked.
Egress	not stated
Identity	national-id national-id: the notebook requires authorising and linking an Alibaba Cloud account, and 'an Alibaba Cloud account requires real-name verification', quoted from a third-party guide rather than a vendor page. The student route is narrower again: a full-time student at a mainland China university, on an account already through individual real-name verification, completed by Alipay scan, valid 30 days and stopping on graduation. No card requirement, hold, debit or prepaid statement appears in any evidence row: unknown.
Limits	retention-clamped the 100 hours are a one-time grant that only ticks while the instance runs, a single boot is capped at about 8 to 10 hours, and only files under /workspace persist, so the 192 GB card is reachable in 8-hour slices; the harder wall is that the ModelScope notebook needs a real-name-verified Alibaba Cloud account linked. source the 100 hours are a one-time grant that only ticks while the instance runs, a single boot is capped at about 8 to 10 hours, and only files under /workspace persist, so the 192 GB card is reachable in 8-hour slices; the harder wall is that the ModelScope notebook needs a real-name-verified Alibaba Cloud account linked.
Binding	retention-clamped the 100 hours are a one-time grant that only ticks while the instance runs, a single boot is capped at about 8 to 10 hours, and only files under /workspace persist, so the 192 GB card is reachable in 8-hour slices; the harder wall is that the ModelScope notebook needs a real-name-verified Alibaba Cloud account linked.
Interface	rest inferred from the tooling the provider documents; not probed
Tooling	not stated
Composes with	192 GB of VRAM is enough to hold a 70B model at half precision with room for activations, which no other free resource in this corpus reaches; it composes with AMD Radeon Cloud's Token Factory for serving and leaves the GPU hours for training.
Reliability	no reliability event recorded
Notes	The corpus record modelscope-studio.json is fails-identity for the Studio tier on the .cn site, with 2 vCPU and 8 GiB, and it explicitly records that the .ai site is a separate lane. ⛔ That record is right about the Studio and misses this entirely: the AMD notebook is a different product with a different number, and 192 GB of VRAM behind a real-name gate is the largest single free GPU measured. The 1,000-hour total is the vendor's own arithmetic over three paths, two of which are content review, so the realistic first number is 100. Whether modelscope.ai carries the same AMD offer is unverified and is the single most valuable open question here.
Evidence rung	reported
Checked	2026-10-22 (live); stale on 2026-11-05
Parsed numbers	`$.limits[0].value` limit:retention-clamped 100.0 hours

Evidence (10)

date	source	what it proved
2026-10-22	www.modelscope.cn/learn/433641	the official guide exists at this URL and its title, read through the broker, reads 'modelscope x AMD developer incentive programme: up to 1000 hours and above of free compute, a collection guide'; the body did not extract, so the figures below rest on the republished text
2026-10-01	zhuanlan.zhihu.com/p/2044198690726860697	a republishing of the ModelScope guide by the AMD China account, quoted: Step 0 pays 100 hours and is the prerequisite for everything else; 25 hours per article capped at 50; 50 hours per notebook capped at 150; 700 hours or more for a deployed application; the three paths are independent and stack
2026-09-20	www.aipromonow.com/promos/modelscope-amd-100h-gpu-notebook-s	reported shape and route: the AMD environment carries a NEW tag and reads about 8 cores, 200 GB RAM and 192 GB of VRAM on a ROCm plus PyTorch image; the 100 hours arrive after binding an AMD developer account from the My Notebook page
2026-10-01	juejin.cn/post/7675623932943548479	a comparison table of the three ModelScope environments from a user's own console: CPU 8 cores 32 GB marked long-term use, NVIDIA GPU 8 cores 32 GB with 24 GB of VRAM on a limited quota, AMD GPU 8 cores 200 GB with 192 GB of VRAM at 100 hours, image ubuntu22.04-rocm7.2.3-py312-torch2.11.0
2026-09-19	linux.do/t/topic/2922778	first-hand reports, translated: 192 GB of VRAM confirmed; 'single boot at most eight hours, and only files inside workspace persist'; one user reports resource shortage preventing an instance from starting; another says a further 200 hours is available by writing articles
2026-10-02	lanmao.ai/o/modelscope-notebook-free-gpu/	the identity gate, quoted: using Notebook requires authorising and linking an Alibaba Cloud account, and an Alibaba Cloud account requires real-name verification
2026-10-22	lanmao.ai/zh/o/modelscope-notebook-free-gpu/	Every new user gets a free initial GPU allowance of 36 hours, a single instance run capped at 10 hours, 100 GB of free persistent storage, and a recorded check date of 2026-10-02; the page cites modelscope.cn/docs/notebook/intro as the vendor source.
2026-10-22	developer.aliyun.com/article/1628727	Alibaba Cloud developer community post headlined 300 hours of free GPU for students with an edu email, linking to modelscope.cn/brand/view/GPU.
2026-10-22	developer.aliyun.com/article/1605205	Binding an Alibaba Cloud account also gets 100 hours of free GPU compute on ModelScope Notebook, the smaller figure that recurs across Alibaba's own posts.
2026-10-22	help.aliyun.com/zh/account/student-identity-verification	The Alibaba Cloud student identity check requires a full-time student at a mainland China university and an account already through individual real-name verification, completed by Alipay scan; verification lasts 30 days and stops working on graduation.

More fields (1)

{
 "wave2": {
  "class": "vendor-developer-program",
  "free_tier_kind": "one-time-credit",
  "shape": {
   "vcpu": 8,
   "ram_gib": 200,
   "gpu_class": "AMD Instinct MI300X",
   "gpu_count": 1,
   "vram_gib": 192
  }
 }
}

Unknown or withheld values (7)

$.egress.metered egress-unknown
$.egress.terms not-stated
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed
$.retention.notes process
$.tooling not-stated

ModelScope Studio Compute

fails: identitycountry-phoneno cardphoneU

modelscope.cn/openapi/v1/studios/hardware · notebook runtime · compute/modelscope-studio · JSON · compare

Free tier	a free Studio hardware tier: platform/2v-cpu-8g-mem on modelscope.cn and platform/2v-cpu-16g-mem on modelscope.ai, both resource_type free with no GPU unit none; metered on wall-clock
Price	none stated
Per-object cap	not applicable
Retention	the docs define a Sleeping state after a long period without access, woken by visiting the URL; the exact timeout is not published
Egress	not published
Identity	country-phone a Chinese phone number is required on modelscope.cn; modelscope.ai is the international site
Limits	retention-clamped a Studio enters a Sleeping state after a long period without access; the timeout is not published source per-object-cap 2 vCPU with 8 GiB or 16 GiB depending on the site source write-gated an account is required and the free quota carries a time limit source the sleeping timeout is a documented absence, so the idle policy cannot be designed against
Binding
Interface	ssh a Gradio, Streamlit, Docker or static SDK on a public *.ms.show URL; Tailscale plus SSH is reported as achievable
Tooling	the Studio SDKs; /mnt/workspace is OSS-backed
Reliability	no reliability event recorded
Notes	The hardware list is a JSON endpoint and the reader extracts HTML, so the free Studio tier cannot be re-read this way. The record's identity finding still stands from the 2026-10-06 read and is not re-asserted here.
Evidence rung	unverified
Checked	2026-10-23 (live); stale on 2026-11-06
Parsed numbers	`$.limits[1].value` limit:per-object-cap 8.0 GiB

Evidence (3)

date	source	what it proved
2026-10-06	modelscope.cn/openapi/v1/studios/hardware	hardware list includes platform/2v-cpu-8g-mem with resource_type free and gradio, streamlit, docker and static SDKs
2026-10-06	modelscope.cn/openapi/v1/studios/hardware	the two sites serve different anonymous free tiers: 2v-cpu-8g-mem on .cn and 2v-cpu-16g-mem on .ai
2026-10-22	modelscope.cn/openapi/v1/studios/hardware	unverified: the model-hardware endpoint returns JSON, and the broker refuses it as 'not html'. Tried broker read, ?__provider=firecrawl, raw fetch: the firecrawl route returned 1,369 bytes, which is the broker's refusal rather than the hardware list. The record's 2-vCPU free tier rests on the 2026-10-06 read

More fields (9)

{
 "shape": null,
 "billing_model": "flat-pool",
 "session_cap": {
  "wall_clock_min": null,
  "idle_reap_s": null,
  "restart_required": false,
  "note": "sleeps after a long period without access; timeout unpublished"
 },
 "keep_rate": null,
 "persistent_storage": {
  "included_gib": null,
  "price_gib_month": "OSS-backed workspace",
  "durability": "/mnt/workspace",
  "survives_stop": true,
  "survives_destroy": false,
  "max_retention": "while the Studio exists"
 },
 "network_out": {
  "allow_outbound": true,
  "allow_inbound_listen": true,
  "ports": "the *.ms.show URL",
  "static_ip": false,
  "ipv6": null,
  "per_session_bytes": null
 },
 "free_tier_kind": "allowance",
 "census_rows": [
  "modelscope-studio"
 ],
 "identity_split": "modelscope.cn requires a Chinese phone; modelscope.ai is the international lane with a different tier"
}

Unknown or withheld values (6)

$.chaining withheld
$.egress.metered egress-unknown
$.free_tier.notes withheld
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed