Technology RadarTechnology Radar
Trial

DeepSeek R1, released January 2025, was the model that proved open-weight chain-of-thought reasoning could compete with proprietary models -- sending shockwaves through the industry and briefly triggering a stock market reaction. Over a year later, it remains one of the most widely deployed open reasoning models.

Why It's in Trial

DeepSeek R1 earned its place through genuine industry impact:

  • Pioneered open-weight reasoning: First open model to match o1-class chain-of-thought capabilities, demonstrating that reasoning is not exclusive to proprietary labs
  • Strong benchmark performance: AIME 2024 79.8%, MATH-500 97.3% -- competitive with models many times its inference cost
  • Distilled variants: The R1-Distill family (1.5B to 70B, built on Qwen and Llama bases) made reasoning accessible on consumer hardware and edge devices
  • MIT license: Unrestricted commercial use, modification, and redistribution
  • Broad ecosystem support: Available via DeepSeek API, Together.ai, Fireworks.ai, OpenRouter, Ollama, and vLLM

It sits in Trial rather than Adopt because:

  • Newer models (DeepSeek V3.1 Terminus, GLM-5, Grok 4.2) have surpassed it on coding and reasoning benchmarks
  • The same data sovereignty concerns as other DeepSeek models apply (see DeepSeek V3 entry)
  • Reasoning traces can be verbose, increasing token costs on long tasks

The "Sputnik Moment"

R1's release in January 2025 was widely described as a "Sputnik moment" for AI. It demonstrated frontier-level reasoning at a fraction of the training cost claimed by Western labs, forced a re-evaluation of AI cost assumptions, and briefly wiped hundreds of billions off NVIDIA's market cap. The open weights enabled massive community research into how chain-of-thought reasoning works.

Architecture

DeepSeek R1 uses the same Mixture of Experts architecture as the V3 line:

  • 671B total parameters, 37B active per token
  • Transparent chain-of-thought visible in outputs (reasoning traces)
  • Trained with reinforcement learning to develop reasoning capabilities

Distilled Variants

Variant Base Model Parameters Use Case
R1-Distill-Qwen-1.5B Qwen 2.5 1.5B Edge / mobile
R1-Distill-Qwen-7B Qwen 2.5 7B Consumer GPU
R1-Distill-Qwen-32B Qwen 2.5 32B Workstation
R1-Distill-Llama-8B Llama 3.1 8B Consumer GPU
R1-Distill-Llama-70B Llama 3.1 70B Data center

Relationship to DeepSeek V3 / V3.1 Terminus

R1 is the reasoning-focused model in DeepSeek's lineup. For general-purpose coding and chat, see the DeepSeek V3.1 Terminus entry. V3.1 Terminus incorporates hybrid thinking/non-thinking modes that subsume much of R1's reasoning capability with better agentic tool use.

Key Characteristics

Property Value
Total parameters 671B (MoE)
Active parameters 37B per token
Context window 128,000 tokens
License MIT
Provider DeepSeek
Release date January 20, 2025
Weights Hugging Face: deepseek-ai/DeepSeek-R1

Further Reading