DeepSeek R1 models
14 DeepSeek R1 models from DeepSeek on Hugging Face, from 1.8B to 684.5B parameters, published by DeepSeek in BF16 and F8_E4M3, with contexts from 128K to 160K tokens. The smallest official model, DeepSeek-R1-0528-Qwen3-8B, needs about 18.3 GB of VRAM at its published precision; the cheapest live fit is RTX 3090 at $0.162/hr.
Part of the DeepSeek series
Pick a size
One row per official DeepSeek R1 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| DeepSeek-R1-0528-Qwen3-8B | 8.2B | 18.3 GB | 9.2 GB | 4.6 GB | RTX 3090 | $0.162/hr | 4.5 GB |
| DeepSeek-R1 | 684.5B | 765 GB | – | 383 GB | No live fit | – | 2.1 GB |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (3)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| DeepSeek-R1-0528-Qwen3-8B | 8.2B | BF16 | 18.3 GB | RTX 3090 | $0.162/hr |
| DeepSeek-R1 | 684.5B | F8_E4M3 | 765 GB | No live fit | – |
| DeepSeek-R1-0528 | 684.5B | F8_E4M3 | 765 GB | No live fit | – |
Quantized versions (2)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| DeepSeek-R1-Distill-Qwen-32B-FP8-dynamic | 32.8B | F8_E4M3 | 36.6 GB | RTX 4090 | $0.781/hr |
| DeepSeek-R1-Distill-Llama-70B-FP8-dynamic | 70.6B | F8_E4M3 | 78.9 GB | RTX PRO 6000 | $1.38/hr |
Fine-tunes and community models (9)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| DeepSeek-R1-Distill-Qwen-1.5B | 1.8B | BF16 | 4.0 GB | RTX 3070 | $0.132/hr |
| DeepSeek-R1-Distill-Qwen-7B | 7.6B | BF16 | 17.0 GB | RTX 3090 | $0.162/hr |
| LoCo-7B | 7.6B | BF16 | 17.0 GB | RTX 3090 | $0.162/hr |
| DeepSeek-R1-Distill-Llama-8B | 8.0B | BF16 | 17.9 GB | RTX 3090 | $0.162/hr |
| DeepSeek-R1-Distill-Llama-8B-abliterated | 8.0B | BF16 | 17.9 GB | RTX 3090 | $0.162/hr |
| DeepSeek-R1-Distill-Qwen-14B | 14.8B | BF16 | 33.0 GB | A100 | $0.462/hr |
| DeepSeek-R1-Distill-Qwen-14B-abliterated-v2 | 14.8B | BF16 | 33.0 GB | A100 | $0.462/hr |
| DeepSeek-R1-Distill-Qwen-32B | 32.8B | BF16 | 73.2 GB | RTX PRO 6000 | $1.38/hr |
| DeepSeek-R1-Distill-Llama-70B | 70.6B | BF16 | 158 GB | RTX 3090 × 7 | $1.13/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
Run and fine-tune DeepSeek R1
- How to run DeepSeek R1: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune DeepSeek R1: LoRA and QLoRA memory per model, and the license limits on derivatives.
More from DeepSeek
- DeepSeek V3 (7 models)
- DeepSeek OCR (2 models)
- DeepSeek Coder (6 models)
- DeepSeek V2 (5 models)
- DeepSeek VL (3 models)
- DeepSeek MoE (2 models)
- DeepSeek V4 (3 models)
