DeepSeek R1 70B
by DeepSeek · deepseek-r1 family
70B
parameters
text-generation code-generation reasoning math creative-writing summarization
DeepSeek R1 70B is the largest distilled reasoning model in the DeepSeek R1 series, based on the Llama 3.3 70B architecture. It captures the most reasoning capability from the full DeepSeek R1 671B model through distillation, delivering exceptional performance on complex reasoning tasks. This model approaches the reasoning quality of the full R1 model on many benchmarks while requiring far less compute. It excels at advanced mathematics, competitive programming, scientific reasoning, and complex analytical tasks. Multi-GPU setups are recommended for comfortable inference.
Quick Start with Ollama
ollama run 70b-q4_K_M | Creator | DeepSeek |
| Parameters | 70B |
| Architecture | transformer-decoder |
| Context | 128K tokens |
| Released | Jan 20, 2025 |
| License | MIT |
| Ollama | deepseek-r1:70b |
Quantization Options
| Format | File Size | VRAM Required | Quality | Ollama Tag |
|---|---|---|---|---|
| Q4_K_M rec | 34.9 GB | 43.5 GB | | 70b-q4_K_M |
| Q5_K_M | 40.8 GB | 50.5 GB | | 70b-q5_K_M |
| Q8_0 | 64.8 GB | 72 GB | | 70b-q8_0 |
Compatible Hardware
Q4_K_M requires 43.5 GB VRAM
Benchmark Scores
85.5
mmlu