Can you own it?

Intended & out-of-scope use

DeepSeek V3 and R1 are large mixture-of-experts language models intended as general assistant, reasoning, and coding engines. V3 is the general-purpose instruct model; R1 is the reasoning-tuned sibling that exposes an explicit chain-of-thought mode and put the lab on the global map. The community R1-Distill checkpoints (Llama and Qwen bases) target the same reasoning/coding tasks on far smaller hardware.

Out-of-scope, in OneHill's read: any use requiring topic-neutral factuality (the models exhibit 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 rather than trusted to self-police. Adopters should also separate the open weights (this entry's subject) from DeepSeek's hosted app/API, which is a different product with its own privacy profile.

Known limitations, bias & failure modes

The distinctive limitation is not hallucination (which is typical for the class) but topic censorship: refusals and steered answers aligned with Chinese content rules on certain political topics, a documented behaviour that affects neutrality and factuality for some use cases. Safety tuning is lighter than at Western frontier labs, so known jailbreaks are easier to elicit and residual prompt-injection susceptibility is present. No first-party guard/classifier model ships. Separately, and 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 sits in the open_weights tier (dimension ceiling 3). The weights are open and MIT-licensed, and the documentation - detailed technical reports - is a genuine strength. But training data is closed, training code is only partially available, and evaluation is partial. That mix meets the open-weights anchor rather than the fully-open top anchor: you get runnable, redistributable weights and strong write-ups, but not a reproducible training recipe or open data.

License terms & what you may do

The V3/R1 weights are released under the OSI-approved MIT license - permissive, with commercial use and even distillation allowed, and no field-of-use restriction or scale threshold. That is a materially more open posture than the community-license families and is the single biggest legal plus of the release. One caveat: some earlier DeepSeek releases used a custom license, so confirm the MIT grant per checkpoint - for the V3/R1 weights it holds.

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, earning a 4). Safetensors matters here: the format is data-only, so loading the canonical weights cannot execute code - the self-hosted open weights are supply-chain-safe to load. Short of a 5 only for the absence of cryptographic signing or provenance attestation. The important caveat is the sprawling ecosystem of third-party distills and quantizations (for example R1-Distill-Llama-70B): these are separate artifacts whose trust equals their uploader, not the canonical org. Pin the revision, verify the checksum, and prefer the canonical org or a host you already trust.

Open weights vs the hosted service - read this. This entry documents the open weights, which you run on your own infrastructure with no data leaving your box. That is distinct from DeepSeek's hosted app and API, which has faced data-privacy scrutiny and outright bans in several jurisdictions and government bodies. Those hosted-service issues are not inherited by locally-run weights, but they are directly relevant if you call DeepSeek's own API, where your prompts leave your control. Self-hosting the safetensors sidesteps that concern entirely.

EU AI Act posture

DeepSeek is a GPAI model, and V3/R1's training compute very likely exceeds the 1e25-FLOP threshold, triggering the systemic-risk presumption. This produces a subtle but important outcome. On the license alone, MIT is a real free-and-open-source license, so V3/R1 would qualify for the Article 53 open-source exemption. But the exemption is void for systemic-risk models - so the full Article 53 + 55 obligations apply, and they are not met: no Article 55 documentation, no copyright policy, no training-content summary. For a downstream EU deployer this is a concrete compliance gap - a systemic-risk GPAI model shipping without the documentation you would need to assemble your own obligations. This is the entry's hard flag, and it caps the grade at C.

Benchmarks & evaluation

On public-leaderboard evidence, R1 is a strong reasoning model and V3 is competitive among large open-weight models. OneHill has not re-run these benchmarks this cycle, so performance is capped at 4 and no specific figures are asserted as verified - the framing here is third-party only. Evaluate against your own task before adoption, particularly where topic-neutral factuality matters.