Can you own it?
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.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Intended & out-of-scope use
DeepSeek-V3 (MIT) is the general-purpose instruct line of DeepSeek-V3 as released under a clean MIT licence - the V3-0324 refresh and the V3.1 hybrid generation, both 671B-total / 37B-active Mixture-of-Experts models with a 128K context. It is intended as a general assistant, coding, and analysis engine. It is a separate entry from the original December-2024 DeepSeek-V3 because the licence changed: original V3 weights shipped under the custom DeepSeek License Agreement; V3-0324 and later moved to MIT.
Out-of-scope, in OneHill's read: any use requiring topic-neutral factuality (documented censorship on certain political topics), and any unguarded customer-facing or high-stakes deployment - the lighter safety tuning means the weights must be wrapped in your own safety system. Separate the open weights from DeepSeek's hosted app/API, a different product with its own privacy profile.
Known limitations, bias & failure modes
The distinctive limitation is topic censorship: refusals and steered answers aligned with Chinese content rules on certain political topics. Safety tuning is lighter than at Western frontier labs, known jailbreaks are easier to elicit, and no first-party guard/classifier model ships. V3.1's hybrid reasoning mode adds an output convention you must handle. Recorded factually: this is a China-origin model; some organisations restrict China-origin models by policy - a governance consideration, not a capability judgment.
Openness tier & components
DeepSeek-V3 (MIT) sits in the open_weights tier (dimension ceiling 3). The weights are
open and MIT-licensed, and the documentation - a detailed technical report - is a
strength. But training data is closed, training code partial, evaluation partial. That mix
meets the open-weights anchor rather than the fully-open top anchor.
License terms & what you may do
The V3-0324 and V3.1 weights are released under the OSI-approved MIT License -
permissive, with commercial use and modification allowed and no field-of-use restriction.
This is a materially more open posture than the original December-2024 V3 weights, which
carry the custom "DeepSeek License Agreement v1.0" with RAIL-style use restrictions (that
generation is the deepseek-v3-original entry). Confirm you are on a V3-0324-or-later
checkpoint to get the MIT grant. The MIT move is what lifts use-and-modify to strong and
ownership to substantial.
Supply-chain provenance
The canonical source is the verified deepseek-ai organisation on Hugging Face,
distributing safetensors with checksums and no malicious-checkpoint incident on the
canonical org (checklist ~5/8). Safetensors is data-only, so loading the canonical weights
cannot execute code. Short of a higher score only for the absence of cryptographic signing.
The caveat is the sprawling ecosystem of third-party quantizations - separate artifacts
whose trust equals their uploader. Pin the revision, verify the checksum, prefer the
canonical org.
Open weights vs the hosted service. This entry documents the open weights, run on your own infrastructure with no data leaving your box - distinct from DeepSeek's hosted app and API, which has faced data-privacy scrutiny and bans in several jurisdictions. Those issues are not inherited by locally-run weights but are relevant if you call DeepSeek's own API. Self-hosting the safetensors sidesteps the concern.
EU AI Act posture
DeepSeek-V3 (MIT) is a GPAI model. The V3 technical report cites ~2.788M H800 GPU-hours over 14.8T tokens; at 671B the systemic-risk question is live, but DeepSeek publishes no FLOP budget, so the 1e25 crossing is not a grounded figure. On the licence alone MIT is a real free-and-open-source licence, a strong Article 53 exemption candidate. But the surviving obligations are unmet: no Article 55 documentation, no copyright policy, no training-content summary, and a China-based provider is unlikely to furnish an EU package. For a downstream EU deployer this is a concrete compliance gap despite the permissive licence - legal scores 2 and the grade holds at C.
Benchmarks & evaluation
On public-leaderboard evidence the MIT V3 generations are competitive among large open-weight instruct models. OneHill has not re-run these benchmarks this cycle, so performance is capped at 4 and no specific figures are asserted as verified. Evaluate against your own task before adoption, particularly where topic-neutral factuality matters.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
StrongThe V3-0324 and V3.1 generations are MIT-licensed (OSI, permissive, ungated), with commercial use and any modifications permitted and no field-of-use limit - a clean permissive grant, so use-and-modify is strong.
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateWeights are inspectable and there is a detailed technical report, but the training data and code are closed, and the weights carry China-aligned topic filtering you cannot inspect - open_weights, so moderate.
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.