DeepSeek-R1-Distill-Llama

TAS Score: S3/3 — D3/5 — A3 / T1 (T1 not T2: the weights carry a contested license — DeepSeek asserts MIT, but the base is Llama-3, which is governed by the Llama Community License (not OSI-approved). The assertion may not hold over the inherited base. This is a Transparency Fragility case: the stated license cannot be relied upon.)

Family: DeepSeek-R1 (full) · R1-Distill-Qwen · R1-Distill-Llama Scores the Llama-based distilled variants (8B / 70B). Technically identical in autonomy to the Qwen distills, but scores lower on transparency because of an inherited license conflict.

Brief Description

Dense reasoning models distilled from DeepSeek-R1 onto the Llama-3 series (Llama-3.1-8B, Llama-3.3-70B). Fine-tuned on 800k samples generated by the full R1. Run on consumer hardware via Ollama, llama.cpp, or vLLM.

Why T1, not T2 — the license is contested. DeepSeek releases these weights under MIT. But the base models are Llama-3, distributed under the Llama Community License, which is not OSI-approved (it carries field-of-use restrictions and a 700M-MAU commercial gate). It is unresolved whether an MIT grant is even valid over Llama-derived weights. TAS does not accept a stated license at face value when the underlying provenance carries a conflicting, non-OSI license — the same rule that caps BSL/Sustainable-Use projects at T1. The autonomy (A3) is unaffected: you can still run, export, and recover the weights freely.

Architectural Role

Compute/inference layer: consumer-hardware reasoning model, Llama-based. Functionally interchangeable with the Qwen distills; the difference is legal provenance, not capability.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts
  • Runs on consumer hardware
  • License is uncontested — MIT is asserted but conflicts with the inherited Llama Community License

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Weights stay on disk.
Exit ✅ Standard model format. Weights run and export freely.
Recoverability ✅ Re-download from HuggingFace or restore from backup.
Visibility ⚠️ Contested license. MIT asserted over a Llama-3 base under the non-OSI Llama Community License. The right to use may not be what the label says.
External Dependencies ✅ Self-hosted: none.

Configuration (Minimal) 

# Via Ollama (Llama-based distills)
ollama pull deepseek-r1:8b        # Llama-3.1-8B base
ollama pull deepseek-r1:70b       # Llama-3.3-70B base

If uncontested licensing matters to you, prefer the Qwen-based distills (1.5b / 7b / 14b / 32b tags), scored T2.

Alternatives

Alternative Autonomy Notes
R1-Distill-Qwen A3 / T2 Same capability, Apache base — clean license.
MiroThinker A3 / T2 Verification-centric reasoning. Apache-2.0.
Ollama + Qwen3 A3 / T2 Simpler reasoning, clean license.

Trajectory

Direction: mixed

Open weights, permissive intent — but the inherited Llama licensing creates a standing ambiguity that the distiller cannot resolve. Provenance flows downstream: the derivative cannot be cleaner than its base.

Signal assessment:

Signal Status Evidence
License ⚠️ MIT asserted over non-OSI Llama Community License — contested.
Feature gating ✅ Open weights, 8B and 70B.
Self-hosting ✅ Ollama, vLLM, llama.cpp compatible.
Governance ⚠️ Parent project: corporate, Chinese regulatory environment; base governed by Meta.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

This site uses Just the Docs, a documentation theme for Jekyll.