fshy

fshy / Storage

OpenCSG 传神社区

opencsg.com · Storage · 1 record · JSON

OpenCSG 传神社区 (CSGHub SaaS, opencsg.com / hub.opencsg.com) Storage

caveatnational-idno cardphoneDretention-clamped

opencsg.com · model-hub · storage/opencsg · JSON · compare

Free tierNo published byte quota anywhere in the docs — and this is deliberate: the upload page states verbatim '上传大小和超时限制由站点、代理和存储配置决定' (upload size and timeout limits are determined by the site, proxy and storage configuration) and explicitly warns '不要把站点限制理解为 Git LFS 协议的统一上限' (do not read a site limit as a universal Git LFS protocol ceiling). What is measurable instead is what they actually host: the canonical deepseek-ai/DeepSeek-V3 repo is 176 files totalling 688.6 GB, of which 163 are shards above 1 GB. That is a single repo at 0.69 TB, i.e. the repo-level ceiling is at minimum ~700 GB and is shard-driven, not quota-driven.
unit objects; metered on stored-bytes
Pricenone stated
Per-object capnot stated
Retention(b) Presigned OSS URLs carry X-Amz-Expires=259200 — 72 hours, so re-resolve per request rather than caching a URL (§1b's expiring-redirect rule) (3 days)
EgressFree and unmetered. Reads are served from opencsg-gitea-lfs-new.oss-cn-beijing.aliyuncs.com (Alibaba Cloud OSS, Beijing) via presigned URLs; no per-GB line, no egress meter, no bandwidth charge appears in any billing doc. The only billing surface is the AI gateway (Token spend), not bytes.
Identitynational-id email + any phone + password. Docs state registration collects 'username, password, phone number, and email'. No national-ID or card requirement documented for the SaaS.
Limits
  • per-ip-shaping 2.70 MB/s source
  • per-ip-shaping 1.16 MB/s source
  • retention-clamped Expires=259200 — 72 hours source
The real limits measured live, not documented: (a) the plain resolve path is NOT the download path — https://hub.opencsg.com/models/Qwen/Qwen3-8B/resolve/main/model-00001-of-00005.safetensors returns HTTP 200 with a 135-byte Git-LFS pointer (Content-Type: text/plain), which is exactly the round-2 'status-code theatre' trap. You must POST the LFS batch endpoint to get bytes. (b) Presigned OSS URLs carry X-Amz-Expires=259200 — 72 hours, so re-resolve per request rather than caching a URL (§1b's expiring-redirect rule). (c) Per-stream shaping is severe: 2.70 MB/s on one connection (16 MiB in 6.2 s) and 1.16 MB/s on a 32 MiB pull; 8 parallel ranged GETs reached 14.49 MB/s and 15.50 MB/s across two separate tests. That is near-linear scaling (5.4x at 8 streams), i.e. it is per-connection shaping, not an aggregate cap — the same shape as Zenodo. (d) No documented upload-bytes/day cap, no retention/expiry on repos, no pull-count limit observed.
Bindingretention-clamped a clock, not a quota: bytes leave whether or not the client is finished with them
Interfacerest, git, rclone, cli git
ToolingThree real clients, all first-party: (1) git + git-lfs against https://opencsg.com/models/<ns>/<repo>.git; (2) csghub-sdk / csghub-cli (pip install csghub-sdk), with csghub-cli upload-large-folder purpose-built for multi-GB model trees; (3) a documented REST API at hub.opencsg.com/api/v1 (models, tree, branches, last_commit all return JSON). The whole platform is open source: github.com/OpenCSGs/csghub (4,071 stars) and csghub-server (1,039 stars), so the server-side quota logic is auditable. No rclone backend.
Composes withRepos are free and unlimited per account, so capacity is N repos x per-repo size with no account total. A 10 TiB hot working set = ~15 repos the size of DeepSeek-V3. Accounts are self-service (username/password + phone + email) and orgs are free, so account multiplication is available on top. Critically, CSGHub ships an XNet storage backend with chunk-level dedup and '秒传' instant-upload, which means re-uploading a dataset that already exists removes bytes entirely — the marginal cost of a fork or a version is near zero.
ReliabilityReads depend on Alibaba OSS presigned URLs, so a URL cached for more than 72 h silently dies (301/0-byte) — resolve-then-fan-out within the window. No observed rate limit on the anonymous LFS batch endpoint. Data does not expire; repos have branches and commit history, so it is mutable rather than an archive.
NotesNo published byte quota anywhere in the docs — and this is deliberate: the upload page states verbatim '上传大小和超时限制由站点、代理和存储配置决定' (upload size and timeout limits are determined by the site, proxy and storage configuration) and explicitly warns '不要把站点限制理解为 Git LFS 协议的统一上限' (do not read a site limit as a universal Git LFS protocol ceiling). What is measurable instead is what they actually host: the canonical deepseek-ai/DeepSeek-V3 repo is 176 files totalling 688.6 GB, of which 163 are shards above 1 GB. That is a single repo at 0.69 TB, i.e. the repo-level ceiling is at minimum ~700 GB and is shard-driven, not quota-driven.
Evidence rungdocumented The object cap and the repository total were recomputed from the live tree API. Was reported; now documented.
Checked2026-10-21 (inherited); stale on 2026-11-20
Parsed numbers
  • $.limits[0].value limit:per-ip-shaping 2.7 MB
  • $.limits[1].value limit:per-ip-shaping 1.16 MB
  • $.limits[2].value limit:retention-clamped 72.0 hours
Evidence (8)
datesourcewhat it proved
2026-10-21hub.opencsg.com/api/v1/models/deepseek-ai/DeepSeek-V3/treeprobe: GET -> 200, 176 files, summed size 688.6 GB, 163 shards >1 GB (largest 5.2 GB)
2026-10-21hub.opencsg.com/models/Qwen/Qwen3-8B.git/info/lfs/objects/baprobe: POST with {operation:download} -> 200, JSON containing actions.download.href to opencsg-gitea-lfs-new.oss-cn-beijing.aliyuncs.com with X-Amz-Expires=259200 — anonymous, no token required
2026-10-21hub.opencsg.com/models/Qwen/Qwen3-8B/resolve/main/model-0000probe: GET -> HTTP 200 but Content-Length 135, body is a git-lfs pointer ('version https://git-lfs.github.com/spec/v1 / oid sha256:...'). The 200 is not a file transfer.
2026-10-21withheldprobe: 1 ranged GET 16 MiB -> 206, 2,701,748 B/s. 8 parallel 8 MiB ranged GETs -> 14.49 MB/s aggregate; repeat on the 5.2 GB DeepSeek-V3 shard -> 15.50 MB/s aggregate
2026-10-21www.opencsg.com/docs/model/upload_modelsread live: '上传大小和超时限制由站点、代理和存储配置决定 ... 不要把站点限制理解为 Git LFS 协议的统一上限' (no published quota; limits are site-config)
2026-10-21www.opencsg.com/docs/csghub/101/introductionread live: 'XNet 智能块加速引擎支持切块去重、秒级增量更新与并行下载 ... 整合 Git Server、Git LFS 与对象存储 OSS'
2026-10-21github.com/OpenCSGs/csghub4,071 stars, self-hostable CSGHub CE (so the same stack can be run on your own box)
2026-10-22probe:https://hub.opencsg.com/api/v1/models/deepseek-ai/DeepSeek-V3/treeGET 2026-10-22: 200, 150,045 B; 176 files, 688,603,604,082 B total, 163 shards over 1 GB, largest 6,584,784,447 B
Unknown or withheld values (5)
  • $.evidence[3].url private-url
  • $.free_tier.notes process
  • $.per_object_cap.notes not-extracted
  • $.per_object_cap.stated not-extracted
  • $.retention.notes not-extracted