DeepSeek-R1-Distill-Qwen

TAS Score: S3/3 — D4/5 — A3 / T2 (D4 not D5: parent project’s hosted API and Chinese regulatory environment give the family a mixed trajectory. The distilled weights themselves are cleanly licensed.)

Family: DeepSeek-R1 (full) · R1-Distill-Qwen · R1-Distill-Llama Scores the Qwen-based distilled variants (1.5B / 7B / 14B / 32B) — a separate object from the full R1 and from the Llama-based distills.

Brief Description

Dense reasoning models distilled from DeepSeek-R1 onto the Qwen-2.5 series (1.5B, 7B, 14B, 32B). Fine-tuned on 800k samples generated by the full R1. Run on consumer hardware via Ollama, llama.cpp, or vLLM. Chain-of-thought reasoning visible in output.

Licensing is clean. The distilled weights are released under MIT, and the Qwen-2.5 base is under Apache 2.0 — both OSI-approved. No inherited license conflict, so transparency is uncontested → T2.

Architectural Role

Compute/inference layer: practical, consumer-hardware reasoning model. The everyday self-hosted way to run R1-quality reasoning without a GPU cluster. Local alternative to cloud reasoning APIs.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts
  • Allows data export — MIT weights, Apache base, standard HuggingFace format
  • Runs on a single consumer GPU (7B/14B) or CPU (1.5B)

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Weights stay on disk.
Exit ✅ Standard model format. MIT + Apache — no use restrictions.
Recoverability ✅ Re-download from HuggingFace or restore from backup.
Visibility ✅ MIT weights, Apache 2.0 base. Both OSI-approved. Methodology published.
External Dependencies ✅ Self-hosted: none.

Configuration (Minimal) 

# Via Ollama (Qwen-based distills — runs on consumer hardware)
ollama pull deepseek-r1:7b        # Qwen2.5-Math-7B base
ollama run deepseek-r1:7b

# Larger Qwen variant
ollama pull deepseek-r1:14b       # Qwen2.5-14B base

Note: in Ollama, the 1.5b / 7b / 14b / 32b tags are Qwen-based. The 8b and 70b tags are Llama-based — see R1-Distill-Llama, scored T1.

Alternatives

Alternative Autonomy Notes
Ollama + Qwen3 A3 / T2 Simpler reasoning. Easier setup.
MiroThinker A3 / T2 Verification-centric reasoning. Apache-2.0.
R1-Distill-Llama A3 / T1 Same idea, Llama base — contested license.

Trajectory

Direction: mixed

The distilled weights are cleanly and permissively licensed (MIT + Apache). Trajectory is marked mixed at the family level because the parent DeepSeek project’s hosted API and governance carry the caveats noted on the full R1 card — not because of anything in these weights.

Signal assessment:

Signal Status Evidence
License ✅ MIT weights, Apache 2.0 base. Both OSI-approved.
Feature gating ✅ Open weights, all Qwen sizes.
Self-hosting ✅ Ollama, vLLM, llama.cpp compatible.
Governance ⚠️ Parent project: corporate, Chinese regulatory environment.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

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