# DeepSeek-R1 Architecture and Economics: How Open-Weights Distillation Cuts Inference Budgets

On January 20, 2025, Chinese AI lab DeepSeek released DeepSeek-R1 under an MIT open-source license. The release proved that frontier-grade reasoning performance, rivaling OpenAI o1 on math, coding, and scientific logic, does not require proprietary closed APIs or hundreds of millions of dollars in compute infrastructure.

Alongside the full 671B Mixture-of-Experts (MoE) model, DeepSeek released six distilled models based on Qwen and Llama from 1.5B to 70B parameters. Developers can now run reasoning models locally on consumer hardware or access hosted APIs at roughly 3% to 5% the cost of OpenAI o1.

By using large-scale reinforcement learning (RL) directly on base models without supervised fine-tuning (SFT) cold starts, DeepSeek mapped an efficient path toward test-time reasoning.

---

## Fast Facts

- **Release Date:** January 20, 2025; weights, technical report, and distilled checkpoints published to Hugging Face and GitHub.
- **License:** Permissive MIT license allowing unrestricted commercial use, fine-tuning, and derivative distillation.
- **API Pricing (DeepSeek Official):** $0.14 per million tokens for cached input, $0.55 per million tokens for unmemoized input, and $2.19 per million output tokens.
- **OpenAI Cost Ratio:** DeepSeek-R1 API is roughly 27 times cheaper for input tokens and 27 times cheaper for output tokens compared to OpenAI o1 ($15.00/M input, $60.00/M output).
- **Architecture:** Mixture-of-Experts (MoE) with 671B total parameters, activating 37B parameters per token across 128 routed experts and 2 shared experts.
- **Key Benchmarks:** 79.8% Pass@1 on AIME 2024 (matching OpenAI o1-1217), 97.3% on MATH-500, and 92.5% on Codeforces percentile.
- **Distilled Lineup:** Open checkpoints for Qwen-2.5 (1.5B, 7B, 14B, 32B) and Llama-3.1 (8B, 70B), fine-tuned directly on 800,000 reasoning trajectories.

---

## Technical detailed review: Reinforcement Learning Without SFT

The core technical contribution of DeepSeek-R1 lies in its training methodology, divided into two distinct research milestones: DeepSeek-R1-Zero and DeepSeek-R1.

### DeepSeek-R1-Zero: Pure RL Discovery

Standard post-training relies on massive instruction tuning before RLHF. DeepSeek tested whether a base model could develop reasoning purely through reinforcement learning without human demonstrations.

The team used Group Relative Policy Optimization (GRPO). Instead of training a separate reward model vulnerable to reward hacking, GRPO scored mathematical accuracy, compiler results, and format compliance directly.

```
+--------------------------------------------------------------------------+
|                  DeepSeek-R1 Multi-Stage Pipeline                       |
+--------------------------------------------------------------------------+
[DeepSeek-V3-Base] 
       │
       ▼  (Large-Scale Pure RL on Math/Code via GRPO)
[DeepSeek-R1-Zero] ───> Discovered self-verification & "Aha moments"
       │
       ▼  (Cold-start data curation: 800k curated trajectories)
[Stage 1 Fine-Tuning]
       │
       ▼  (Multi-Task RL: Accuracy + Human-preference alignment)
[DeepSeek-R1 (671B MoE)]
       │
       ▼  (800k reasoning outputs extracted directly)
[Distillation Pass] ───> Qwen-1.5B/7B/14B/32B & Llama-8B/70B
+--------------------------------------------------------------------------+
```

Without human intervention, R1-Zero naturally developed test-time compute scaling: it allocated more tokens to difficult problems, executed internal verification loops, reconsidered flawed assumptions, and exhibited emergent problem-solving behaviors.

### DeepSeek-R1: Fixing Readability and Language Mixing

While R1-Zero proved accurate, it produced chaotic thinking traces with frequent language mixing. To build production DeepSeek-R1, engineers:

1. Seeded the base model with thousands of structured, human-readable cold-start chain-of-thought samples.
2. Ran secondary RL targeting multi-objective rewards: math correctness, test-case passage, and language consistency.
3. Extracted 800,000 reasoning trajectories to fine-tune dense open-source architectures.

### Distillation vs. RL on Dense Models

DeepSeek proved that distilling reasoning traces from a 671B teacher model into smaller dense architectures yields substantially higher accuracy than training those small models through RL from scratch.

 | Model Checkpoint | Architecture / Params | AIME 2024 (Pass@1) | MATH-500 | Codeforces Rating |
|---|---|---|---|---|
| **OpenAI o1-mini** | Proprietary Dense | 63.6% | 90.0% | 1820 |
| **DeepSeek-R1-Distill-Qwen-14B** | Dense 14B | 69.7% | 93.9% | 1481 |
| **DeepSeek-R1-Distill-Qwen-32B** | Dense 32B | 72.6% | 94.3% | 1692 |
| **DeepSeek-R1-Distill-Llama-70B** | Dense 70B | 70.0% | 94.5% | 1633 |
| **DeepSeek-R1 (Full)** | MoE 671B (37B active) | **79.8%** | **97.3%** | **2029** |
| **OpenAI o1 (Full)** | Proprietary | 79.2% | 96.4% | 1873 |

---

## Real-World Utility & Limitations

### Where DeepSeek-R1 Solves Production Bottlenecks

1. **Synthetic Data Generation Pipelines:** Teams can run DeepSeek-R1 to label, verify, and generate thousands of structured reasoning examples to train proprietary in-house models without violating restrictive commercial terms.
2. **Local Edge Reasoning:** The 14B and 32B distilled models fit easily onto single workstation GPUs (e.g., NVIDIA RTX 4090 or Apple M-series chips with 32GB+ Unified Memory), providing air-gapped reasoning for sensitive legal and medical data.
3. **High-Volume Code Evaluation:** The $0.55/$2.19 per million token pricing enables continuous test generation and syntax checks at scales where OpenAI o1 costs would be prohibitive.

### Operational Challenges & Infrastructure Realities

- **Massive Self-Hosting Footprint (671B MoE):** Running the full 671B model requires an 8x H100/H200 cluster (670GB+ VRAM). FP8 quantized deployments still need 320GB+ VRAM, making self-hosting unviable for small teams.
- **Context Drift:** On open-ended prompts lacking deterministic answers, R1 can loop redundantly, inflating output usage.
- **API Rate Limits:** Official cloud endpoints faced severe throttling post-launch. Production systems require multi-provider failovers.

---

**Learn More:** [Claude 3.7 Sonnet Hybrid Reasoning](https://www.usefulainews.com/claude-sonnet-hybrid-reasoning/) →

**Learn More:** [DeepSeek-V3 MoE Inference Scaling](https://www.usefulainews.com/deepseek-moe-inference-scaling/) →

**Learn More:** [Meta Llama Open Weights Guide](https://www.usefulainews.com/meta-llama-open-weights-guide/) →

## Actionable Takeaways

1. **Deploy Distilled 14B/32B for Local Pipelines:** For local offline development, run `ollama run deepseek-r1:14b` or `deepseek-r1:32b`. The 14B model delivers 69.7% on AIME 2024 while running smoothly on a single GPU or Mac Studio.
2. **Benchmark Hosted Endpoints with SGLang/vLLM:** If hosting the full 671B model, deploy using SGLang or vLLM with FP8 quantization and Multi-Head Latent Attention (MLA) optimizations enabled to maintain 30+ tokens per second per GPU node.
3. **Configure Upstream Provider Fallbacks:** If using DeepSeek’s hosted API, configure an automated fallback router (via LiteLLM or OpenRouter) pointing to OpenAI o3-mini or Groq-hosted checkpoints when API latency exceeds 5,000ms.
4. **Strip Chain-of-Thought for Clean UI Delivery:** When serving user-facing applications, parse out the `<think>...</think>` tags using streaming regex filters unless your end-users explicitly demand auditability.