Alpha
Contact
Assess · DeepSeek-V3

Can you own it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 64.4/100

This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess 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 + Reliability
  • UseReliability
  • SupportTransparency

Intended & out-of-scope use

The original DeepSeek-V3 (December 2024) is the 671B-total / 37B-active Mixture-of-Experts Base and Chat model, with 128K context, that established the V3 line. It is intended as a general assistant (Chat) or a foundation for further training (Base). It has its own entry, separate from the later MIT-licensed V3 generations, because its weights fall under the custom "DeepSeek License Agreement v1.0", not MIT.

Out of scope are any use that violates the licence's field-of-use restrictions (see License), any use requiring topic-neutral factuality (documented political censorship), the Base variant in any customer-facing role (it is not safety-tuned), and any unguarded high-stakes deployment. For new work, prefer the MIT V3 generations (deepseek-v3-mit) unless this checkpoint is specifically needed.

Known limitations, bias & failure modes

The model hallucinates like any LLM. Its distinctive limitations are topic censorship aligned with Chinese content rules, lighter safety tuning than Western frontier labs, no first-party guard model, and a use-restricted licence that limits what it may be deployed for. The Base variant is a raw completion model with no safety tuning. It is a China-origin model, which some organisations restrict by policy.

Openness tier & components

The original V3 is in the open_weights tier (dimension ceiling 3). The weights are openly downloadable (ungated) and a detailed technical report documents the model. The licence is conditional (use-restricted), training data is closed and training code is partial. The restrictions bear on legal and ownership, and the tier holds because download is ungated and the grant is irrevocable.

License terms & what you may do

The code repository is MIT. The weights are governed by the "DeepSeek License Agreement, Version 1.0" (non-OSI). The grant is "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable", commercial use is allowed, and "DeepSeek claims no rights in the Output You generate".

The grant is conditioned on RAIL-style use restrictions (Section 5 / Attachment A): no use violating applicable law, no military application, no harm to minors, no generating verifiably false content to harm others, no unauthorised PII distribution, no defamation or harassment, no discrimination, and no automated decisions affecting legal rights. These flow down to derivatives. The patent licence terminates if the licensee sues DeepSeek over the work. The field-of-use restrictions hold use-and-modify at moderate and ownership at partial, one step below the MIT V3 generations.

Supply-chain provenance

The canonical source is the verified deepseek-ai organisation on Hugging Face. It distributes safetensors with checksums and has no malicious-checkpoint incident (checklist ~5/8). Safetensors is data-only, so loading the canonical weights cannot execute code. The score is held back only by the absence of cryptographic signing. Third-party quantizations are separate artifacts whose trust equals their uploader, and any redistribution must carry the DeepSeek License restrictions, not MIT. Pin the revision, verify the checksum and prefer the canonical org.

This entry covers the open weights run on your own infrastructure. DeepSeek's hosted app and API are a separate matter: they have faced data-privacy scrutiny and bans in several jurisdictions. Those issues do not apply to locally run weights but do apply to calls to DeepSeek's own API.

EU AI Act posture

The original V3 is a GPAI model. Unlike the MIT generations, its weight licence is not free-and-open-source (it carries field-of-use restrictions), so the Article 53 open-source exemption does not apply on the licence axis. The surviving obligations are also unmet: no Article 55 documentation, no copyright policy, no training-content summary, and a closed corpus.

The technical report cites ~2.788M H800 GPU-hours. At 671B the systemic-risk question is open, because no FLOP figure is disclosed. For a downstream EU deployer this is a compliance gap, and the grade holds at C.

Benchmarks & evaluation

On public-leaderboard evidence the original V3 was competitive among large open-weight models at release. OneHill has not reproduced the benchmarks, so performance is capped at 4 and no specific figures are asserted as verified. Evaluate against your own task, particularly where topic-neutral factuality matters.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

1

Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?

Moderate

The weight grant is irrevocable and royalty-free and permits commercial use and modification. It is a non-OSI custom licence with binding RAIL-style field-of-use restrictions (no military or unlawful use, no harm to minors, no automated legal-rights decisions, etc.) that flow down to derivatives, so use-and-modify is moderate, not strong. This is the main difference from the MIT V3 generations.

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyThe weights are openly downloadable (ungated) with an open technical report.
Legal2/5how permissive and clean the licence is for real commercial useThe weight grant is irrevocable and permits commercial use, but the licence is a non-OSI custom licence with RAIL-style field-of-use restrictions.
2

TransparencyDo you know what it is: weights, training, behaviour, and legible terms?

Moderate

Weights are inspectable and there is a detailed technical report, but training data and code are closed and the weights carry China-aligned topic filtering that cannot be inspected. Open weights score moderate.

How this scores (AOI sub-dimensions)
Provenance3/5how well we can trace and verify what went into the modelThe deepseek-ai org on Hugging Face is verified, distributes safetensors with checksums, has a clear canonical source and has no malicious-checkpoint incident (checklist ~5/8).
Governance3/5how accountable and well-documented the publisher isDeepSeek is an active, named publisher with a verified org and a track record of technical reports, which meets the score-3 anchor.
What this means for adoptionYou partially own self-hosted original DeepSeek-V3. Self-hosting keeps your data yours under an irrevocable grant (data-control strong, reliability strong). The DeepSeek License Agreement imposes RAIL-style field-of-use restrictions that flow down to derivatives and hold use-and-modify at moderate, so ownership is partial, one step below the MIT V3 generations (deepseek-v3-mit), which reach substantial. For new work prefer the MIT generations unless this checkpoint is specifically needed. Where it is used, honour the use restrictions, deploy behind your own guardrails and treat EU high-stakes use as needing a self-assembled compliance package, since no open-source exemption applies on the licence axis.

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.

Licenceread2026-11-19
The DeepSeek-V3 LICENSE-CODE states: "This code repository is licensed under the MIT License.
Model cardread2026-11-19
DeepSeek-V3 model card on the verified deepseek-ai HF org: 671B total / 37B active MoE, 128K context, safetensors, Base + Chat variants; first-class serving (vLLM, SGLang, llama.cpp, Ollama) and an extensive community quant ecosystem.
Third-party analysisunverified2026-08-03
On independent public leaderboards the original DeepSeek-V3 was competitive among large open-weight models at release (coding, maths, analysis).
Third-party analysisunverified2026-08-03
Independent analysis notes DeepSeek open-weight models apply China-aligned content filtering on politically sensitive topics, with lighter safety tuning than Western frontier labs and no companion guard model.
Third-party analysisunverified2026-08-03
The DeepSeek License Agreement imposes field-of-use restrictions (non-FOSS), and no public EU AI Act training-content summary, copyright policy, or GPAI documentation package is published for DeepSeek-V3; the training corpus is not released.