DeepSeek R1
Developed by DeepSeek
Key Capabilities
- Step-by-step reasoning and chain-of-thought problem solving
- Matches o1-level performance on math and coding benchmarks
- Self-verification and error correction during reasoning
- Distilled versions available for smaller hardware
- MIT license for unrestricted commercial use
VRAM Requirements by Quantization
Choose the right GPU based on your performance and quality needs.
| Model / Quantization | VRAM Required |
|---|---|
| 7B distill FP16 | 14GB |
| 32B distill FP16 | 64GB |
| 70B distill FP16 | 140GB |
| 671B full FP16 | 1.3TB+ |
Use Cases
DeepSeek R1 (671B total (37B active via MoE), distilled variants: 1.5B, 7B, 8B, 14B, 32B, 70B) can be deployed for enterprise AI applications including document processing, code generation, data analysis, and conversational AI. License: MIT License.
Run DeepSeek R1 with Petronella
PTG deploys DeepSeek R1 for enterprises needing o1-class reasoning without cloud API costs. Distilled variants run on a single GPU; full model on DGX infrastructure. MIT licensed, no strings attached.
Recommended Hardware
| Model Size | Recommended GPU |
|---|---|
| 7B distill | RTX 5080 (16GB) |
| 32B distill | RTX 5090 (32GB) or RTX PRO 5000 (48GB) |
| 70B distill | RTX PRO 6000 Blackwell (96GB) |
| 671B full | DGX B300 or multi-node cluster |
Deploy DeepSeek R1 On-Premises
Our team builds GPU-accelerated systems configured and optimized for DeepSeek R1. Private, secure, and fully under your control.