Alpha
Contact
Assess · DeepSeek-R1-Distill

Can you own it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 61.6/100

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

Two DeepSeek-R1-Distill checkpoints are built on Meta Llama bases: Llama-8B (Llama 3.1) and Llama-70B (Llama 3.3). Both are dense reasoning models distilled from DeepSeek-R1. They suit cost-sensitive reasoning and serve as fine-tuning bases within the Llama ecosystem.

Out of scope: any unguarded customer-facing role (the models have no safety tuning of their own), any use that needs topic-neutral factuality (China-aligned filtering carried over from R1), and any use that breaches Meta's Acceptable Use Policy or the 700M-MAU clause.

Known limitations, bias & failure modes

The Llama Community Licence is more restrictive than the Apache-2.0 licence of the Qwen-base distils, and the models have no safety tuning of their own. They carry R1's China-aligned topic censorship, the distillation data is closed, and capability is strong for the size but below frontier (the 8B is modest). They emit explicit reasoning (think-tag) traces.

Openness tier & components

The tier is open_weights (dimension ceiling 3) with a conditional (Llama community) licence. Weights and documentation are open, the distillation data is closed and training code is partial. The restrictions are scored under legal and ownership.

License terms & what you may do

The distils carry the Llama Community Licence (8B on Llama 3.1, 70B on Llama 3.3). It is non-OSI and more restrictive than the Apache-2.0 licence of the Qwen distils. Commercial use is allowed, subject to Meta's Acceptable Use Policy, the "700 million monthly active users" clause (above it a separate Meta licence is required) and the naming and attribution requirements ("Built with Llama"; derivative model names must include "Llama").

These terms pass to derivatives. They hold use-and-modify at moderate and ownership at partial, one step below the Qwen distils.

Supply-chain provenance

The canonical source is the verified deepseek-ai organisation on Hugging Face. Weights ship as safetensors with checksums, and there is no malicious-checkpoint incident (checklist ~5/8). The models are widely mirrored and quantized, and each mirror is a separate artifact whose trust equals its uploader. Redistribution must carry the Llama licence and AUP. Pin the revision, verify the checksum and prefer the canonical org.

EU AI Act posture

The models are GPAI but not systemic-risk (8B-70B). The Llama Community Licence is not FOSS (it carries acceptable-use and 700M-MAU restrictions), so the Article 53 open-source exemption does not apply on the licence axis, unlike the Apache-2.0 Qwen distils. The systemic-risk duties do not arise. The transparency obligations are not exemption-covered, and DeepSeek publishes no copyright policy or training-content summary. Deployers must also honour Meta's AUP and the 700M-MAU clause. Legal scores 2.

Benchmarks & evaluation

The 70B distil is a strong public result for its size and the 8B is modest. OneHill has not reproduced the benchmarks, so no specific figures are asserted as verified. Performance scores 3: strong for the size, below frontier.

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 Llama Community Licence permits commercial use and modification. It is non-OSI, with an Acceptable Use Policy, the 700M-MAU clause (a separate Meta licence above that threshold) and naming and attribution duties ("Built with Llama"; derivative names must include "Llama") that pass to derivatives. Use-and-modify is therefore moderate, below the Apache-2.0 Qwen distils.

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyThe weights are downloadable and documented, which meets the open-weights anchor (3).
Legal2/5how permissive and clean the licence is for real commercial useThe Llama Community Licence is non-OSI, with acceptable-use and 700M-MAU field-of-use restrictions and naming and attribution duties.
2

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

Moderate

Weights are inspectable. 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 modelThe verified deepseek-ai org on Hugging Face is the canonical source.
Governance3/5how accountable and well-documented the publisher isDeepSeek is an active, named publisher with a verified org and technical reports, which meets the score-3 anchor.
What this means for adoptionYou partially own these self-hosted Llama-base R1-Distill models. Self-hosting keeps your data yours (data-control strong, reliability strong). The Llama Community Licence's acceptable-use policy, 700M-MAU clause and naming and attribution duties hold use-and-modify at moderate. Ownership is therefore partial, one step below the Apache-2.0 Qwen distils (deepseek-r1-distill-qwen), which reach substantial. Prefer the Qwen distils unless the Llama toolchain is needed. Where these models are used, honour the Llama terms and deploy behind your own guardrails. No EU 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-R1 README and the distil checkpoint cards state that the R1-Distill-Llama checkpoints are built on Meta Llama bases (Llama-8B on Llama 3.1, Llama-70B on Llama 3.3).
Model cardread2026-11-19
The R1-Distill-Llama model cards on the verified deepseek-ai HF org describe dense 8B and 70B models distilled from R1, with 128K context and safetensors weights.
Third-party analysisunverified2026-08-03
On public leaderboards the R1-Distill-Llama-70B is a strong reasoning model for its size and the 8B 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-Llama checkpoints are small (8B-70B) and below the systemic-risk threshold.