Alpha
Contact
Assess · DeepSeek-R1-Distill

Can you own it?

Ownership levelSubstantialnone·limited·partial·substantial·fullAnalytical input C ยท 64.8/100

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 DeepSeek-R1-Distill checkpoints built on Qwen2.5 bases (1.5B, 7B, 14B, 32B) are dense reasoning models distilled from DeepSeek-R1 into small, single-GPU weights. They are intended for on-device and cost-sensitive reasoning and as cheap fine-tuning bases. They form a separate entry from the flagship R1 and from the Llama-base distils because they are small dense models on Qwen2.5 bases. DeepSeek releases them under the MIT licence.

Out of scope are any unguarded customer-facing role (these are research distils with no safety tuning of their own) and any use that needs topic-neutral factuality (they carry R1's China-aligned filtering). The 1.5B is modest and is not a frontier model.

Known limitations, bias & failure modes

These are research distils with no safety tuning of their own, so deploy them behind your own safety system. They carry R1's China-aligned topic censorship and the distillation data is closed. Capability is strong for the size but not frontier-absolute, and the 1.5B is weakest. They emit explicit reasoning (think-tag) traces that the client must parse.

Openness tier & components

Open-weights tier (dimension ceiling 3). The weights are MIT-licensed and the documentation is open, but the distillation data is closed and the training code is partial. This meets the open-weights anchor.

License terms & what you may do

The distil repositories carry DeepSeek's MIT licence, and the card metadata declares mit. It is OSI-approved and permissive, with no field-of-use restriction. That is more permissive than the Llama Community Licence of the Llama-base distils (deepseek-r1-distill-llama) and the same as the flagship R1. It makes use-and-modify strong and ownership substantial. The Qwen2.5 base weights are Apache-2.0 per their cards, and the DeepSeek card describes the distils as derived from them. Confirm the licence on each checkpoint card.

Supply-chain provenance

The canonical source is the verified deepseek-ai organisation on Hugging Face, with safetensors and checksums and no malicious-checkpoint incident (checklist ~5/8). These are the most widely mirrored and quantized DeepSeek artifacts because they run on consumer hardware. Each mirror is a separate artifact whose trust equals its uploader. Pin the revision, verify the checksum and prefer the canonical org.

EU AI Act posture

These are GPAI models but not systemic-risk: at 1.5B-32B they are orders of magnitude below the 1e25-FLOP presumption. The MIT licence is free and open-source, so the Article 53 open-source exemption applies to the transparency obligations. That makes these the most EU-comfortable DeepSeek artifacts. DeepSeek still publishes no copyright policy and no public training-content summary, the obligations that survive the exemption for all GPAI. Legal scores 3.

Benchmarks & evaluation

The 14B and 32B distils are strong for their size on public benchmarks; the 1.5B is modest. OneHill has not re-run the benchmarks, so performance scores 3 (strong for size, not frontier-absolute) and no specific figure is asserted as verified.

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?

Strong

MIT (OSI-approved, permissive, ungated), granted by DeepSeek over Apache-2.0 Qwen2.5 base weights. Commercial use and any modification are permitted with no field-of-use limit, so use-and-modify is strong, a step above the Llama-base distils.

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyOpen-weights tier: MIT-licensed weights and open documentation, but the distillation data is closed and training code and evaluation are partial.
Legal3/5how permissive and clean the licence is for real commercial useStronger than the 671B family models because, at this size, these are not systemic-risk, so the open-source exemption covers the transparency duties.
2

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

Moderate

Weights are inspectable under the MIT licence, but the distillation data and code are closed and the weights carry R1's China-aligned topic filtering, which cannot be inspected. Open-weights tier, so moderate.

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 with checksums, a clear canonical source and no malicious-checkpoint incident (checklist ~5/8).
Governance3/5how accountable and well-documented the publisher isActive, named publisher (DeepSeek) with a verified org and technical reports, meeting the score-3 anchor.
What this means for adoptionYou substantially own these self-hosted MIT-licensed R1-Distill (Qwen) models. The permissive licence makes use-and-modify strong, and self-hosting, including on-device at the small sizes, keeps your data yours, so data-control is strong too. Ownership stops short of full because the distillation data and code are closed (not reproducible) and the weights carry R1's China-aligned filtering and no safety tuning of their own. Deploy behind your own guardrails.

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 R1-Distill-Qwen repositories (1.5B, 7B, 14B, 32B) carry DeepSeek's MIT LICENSE, and the card metadata declares license mit.
Model cardread2026-11-19
The model cards on the verified deepseek-ai Hugging Face org describe dense 1.5B, 7B, 14B and 32B models distilled from R1, with a 128K context (from the Qwen2.5 bases) and safetensors weights.
Third-party analysisunverified2026-08-03
On public leaderboards the R1-Distill-Qwen-14B and 32B are strong reasoning models for their size; the 1.5B is modest.
Third-party analysisunverified2026-08-03
The R1-Distill checkpoints are research distils with no safety tuning of their own.
Third-party analysisunverified2026-08-03
The R1-Distill-Qwen checkpoints (1.5B-32B) are well below the systemic-risk threshold.