How to deploy and fine-tune DeepSeek models on AWS

Source: HuggingFace Date: 2025-01-30 URL: https://huggingface.co/blog/deepseek-r1-aws

Summary

Integration tutorial covering four deployment paths for DeepSeek-R1 distilled models on AWS: HF Inference Endpoints ($8.30/hr for quantized), Amazon Bedrock Marketplace, SageMaker AI, and EC2 Neuron DL AMI. Hardware recommendations span ml.g6.2xlarge (8B, 1 GPU) to ml.g6.48xlarge (70B, 8 GPUs). Fine-tuning and full-model DeepSeek-R1 (non-distilled) SageMaker support marked as coming soon at time of writing.

Implications

Thread: open-weights ecosystem health / HF as open-source ML hub. The post is a reaction to the DeepSeek-R1 moment — HF and AWS jointly publishing a deployment guide within days of the release shows how rapidly cloud providers move to capture open-weight model deployments. The four-path options (HF Endpoints vs Bedrock vs SageMaker vs direct EC2) reflect genuinely different cost/control tradeoffs. The $8.30/hr entry point for a hosted reasoning model quantized is a reference price worth tracking as the market evolves. Fine-tuning support gaps at launch are expected for newly-released distilled models.