Support · DeepSeek
Will it last?
Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 64.4/100
This page is a projection of the one entry record, the Transparency factor that Support covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
Which domain expands which factor
- AssessUse & modify + Transparency
- ImplementData control + Doesn't fail you
- UseDoesn't fail you
- SupportTransparency
Common problems & fixes
Two classes of issue dominate. First, reasoning-token / chat-template handling: R1
emits inline <think> spans, so blank or malformed answers usually mean the template
wasn't applied or the reasoning span wasn't parsed/stripped - always go through
apply_chat_template and handle the reasoning tags. Second, large-model serving: the
full V3/R1 needs multi-GPU sharding, careful KV-cache/memory planning, and often a
quantized distill instead of the full weights on constrained hardware. If a checkpoint
misbehaves unexpectedly, verify you pulled the canonical org and the pinned revision rather
than a third-party quant.
Versions & changelog
DeepSeek ships versioned families - V3, its point releases (e.g. V3.1) and successors,
alongside the R1 reasoning line and the R1-Distill checkpoints - each as its own Hugging
Face repository under the deepseek-ai org, with commit revisions serving as the
changelog. Pin a specific revision to stay reproducible and watch the org for new
checkpoints; there is no separate SemVer-style release feed for the weights.
Community support runs through Hugging Face model-repo discussions, DeepSeek's
GitHub, and general community forums. There is no first-party enterprise support
contract for the open weights - plan to rely on community channels and your own
operations team.
Tracked known issues
The two distinctive, well-attested known issues are topic censorship on certain
political topics (aligned with Chinese content rules) and the lighter-safety-tuning
caveats - easier jailbreaks and residual prompt-injection susceptibility - documented in
the model card and third-party reporting. Both are behavioural properties to design
controls around, not bugs that a patch removes. Note again that the hosted app/API's
privacy issues are a separate matter from the self-hosted weights.
How this scores
The ownership factor this domain covers, drawn from the one entry record.
2
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
WeakYou hold the weights but cannot see into them - China-aligned topic censorship is baked into behaviour, the training data is unreleased and code only partial, and misuse is unbenchmarked with no third-party red-team.
How this scores (AOI sub-dimensions)
Provenance3/5how well we can trace and verify what went into the modelVerified deepseek-ai org on Hugging Face, safetensors distribution with checksums, clear canonical source and no malicious-checkpoint incident on the canonical org (checklist ~5/8).
Governance3/5how accountable and well-documented the publisher isActive, named publisher with a verified org and a track record of technical reports and releases, meeting the score-3 anchor.
What this means for adoptionYou can own the R1 weights outright under MIT - run, redistribute and distil them self-hosted, and your data stays on your infrastructure - but the family is not uniformly MIT: V3's weights fall under the separate DeepSeek License Agreement with real use restrictions (military, minors, PII, discrimination, legal-rights automation), so check the per-checkpoint licence before assuming a clean grant. Transparency is the deeper cap that holds ownership at partial: China-aligned censorship is baked in and the training is opaque, so you cannot fully know what you are running. Adopt with eyes open for reasoning and distillation where that is acceptable, keep sensitive data off it, and avoid the hosted service - its Privacy Policy stores data on servers in the People's Republic of China and its Terms put disputes under PRC law.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.
Licenceread2026-07-25
DeepSeek-R1, read verbatim: "This code repository and the model weights are licensed under the MIT License.
Licenceread2026-07-25
DeepSeek-V3 licensing splits code from weights (differs from R1): the code repository is MIT ("Copyright (c) 2023 DeepSeek"), but "The use of DeepSeek-V3 Base/Chat models is subject to the Model License" - the separate "DeepSeek License Agreement, Version 1.0", which is NOT MIT.
Third-party analysisread2026-07-25
DeepSeek variants, MIT weights, hosted-service data-privacy scrutiny and bans, topic censorship aligned with Chinese content rules, and China-origin context.
Model cardread2026-07-25
Canonical, verified deepseek-ai organisation on Hugging Face distributing safetensors with checksums.
Third-party analysisread2026-07-25
DeepSeek-R1 is a strong reasoning model and V3 competitive among large open-weight models on public leaderboards.
Privacy Policyread2026-07-25
DeepSeek hosted-service Privacy Policy, read verbatim: "We store the information we collect in secure servers located in the People's Republic of China." Information is retained "as long as necessary to provide our Services".
Terms of serviceread2026-07-25
DeepSeek hosted-service Terms of Use, read verbatim (Section 9.1 Governing Law): the terms are "governed by the laws of the People's Republic of China in the mainland," with negotiation followed by PRC-court litigation and no arbitration.
Technical_reportread2026-07-25
DeepSeek-V3 Technical Report (arXiv 2412.19437): MoE 671B total / 37B active, MLA + DeepSeekMoE, auxiliary-loss-free load balancing, trained on 14.8T tokens using 2.788M H800 GPU-hours.
Technical_reportread2026-07-25
DeepSeek-R1 Technical Report (arXiv 2501.12948), "Incentivizing Reasoning Capability in LLMs via Reinforcement Learning": reasoning incentivized via pure RL with no human-labelled reasoning trajectories, then distilled into smaller models.