DeepSeek-R1

TAS Score: S3/3 — D4/5 — A3 / T2 (D4 not D5: hosted API (api.deepseek.com) collects usage data and requires account — trajectory mixed due to Chinese regulatory environment.)

Family: DeepSeek-R1 (full) · R1-Distill-Qwen · R1-Distill-Llama This card scores the full 671B model only. The distilled variants are separate objects with different base licenses — see the family links above.

⚠️ Two-mode tool. Self-hosted: A3/T2. Hosted API (api.deepseek.com): A1/T1. This card scores the self-hosted open-weight variant.

Brief Description

Full-size open-weight reasoning model from DeepSeek — 671B parameters (Mixture-of-Experts). Trained on DeepSeek-V3-Base (DeepSeek’s own base model). Chain-of-thought reasoning is visible in output. MIT licensed — weights and base are both under a permissive, OSI-approved license, so transparency is uncontested.

Running the full model requires server-class hardware (multi-GPU cluster). For consumer-hardware deployment, use the distilled variants — but note they are different models with different base licenses, scored on their own cards.

Architectural Role

Compute/inference layer: frontier-class reasoning model for complex analysis, run on your own infrastructure. Local alternative to cloud reasoning APIs when you have the hardware to host the full model.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — MIT licensed, standard HuggingFace format
  • Hosted API requires account and sends data to DeepSeek servers

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Model weights stay on disk.
Exit ✅ Standard model format. MIT licensed — no restrictions on use.
Recoverability ✅ Re-download from HuggingFace or restore from backup.
Visibility ✅ MIT license (weights + DeepSeek-V3 base). Open weights. Training methodology published.
External Dependencies ⚠️ Self-hosted: none. Hosted API: DeepSeek infrastructure (China-based).

Configuration (Minimal) 

# Full 671B model — requires a multi-GPU cluster
vllm serve deepseek-ai/DeepSeek-R1 --tensor-parallel-size 8

# Or via SGLang
python -m sglang.launch_server --model deepseek-ai/DeepSeek-R1 --tp 8

For consumer hardware, see R1-Distill-Qwen (Apache base) or R1-Distill-Llama (Llama base — lower transparency).

Alternatives

Alternative Autonomy Notes
MiroThinker A3 / T2 Verification-centric reasoning. Apache-2.0.
R1-Distill-Qwen A3 / T2 Same reasoning, consumer hardware, Apache base.
OpenAI o1 A0 / T0 Cloud-only. Proprietary.

Trajectory

Direction: mixed

MIT licensed — maximally permissive. Groundbreaking open release that challenged frontier labs. But: DeepSeek is a Chinese company subject to regulatory environment. Hosted API subject to content filtering. Future open releases not guaranteed.

Signal assessment:

Signal Status Evidence
License ✅ MIT. Maximally permissive.
Feature gating ✅ Open weights for all variants including full 671B.
Self-hosting ✅ Standard formats. vLLM, SGLang compatible.
Governance ⚠️ Corporate (DeepSeek/High-Flyer). Chinese regulatory environment.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

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