DeepSeek R1 deployment on AWS now costs just $8.30/hour via Hugging Face endpoints

DeepSeek dropped a bomb on the AI world last week with R1, an open-source reasoning model that goes toe-to-toe with OpenAI’s o1 on math, coding, and logic tasks. The model family includes R1-Zero, R1 itself, and six smaller distilled variants built on Llama and Qwen architectures — all now available on Hugging Face.

Hugging Face and AWS are making deployment unusually straightforward. You can spin up any of the six distilled models — or a quantized version of full R1 from Unsloth — directly through Hugging Face Inference Endpoints. The pricing is genuinely eyebrow-raising: full DeepSeek R1 runs at $8.30 per hour on dedicated AWS infrastructure, complete with autoscaling and scale-to-zero capabilities. For developers who’ve been priced out of OpenAI’s API, this changes the math considerably.

The team is also working on Inferentia support, though that’s still in the pipeline. In the meantime, you can deploy the distilled models through Amazon Bedrock Marketplace or SageMaker AI using Hugging Face’s LLM DLCs. The DeepSeek-R1-Distill-Llama-70B variant, for instance, recommends an ml.g6.48xlarge instance with 8 GPUs. JumpStart offers a console-based path, while the Python SageMaker SDK gives you programmatic control.

What makes R1 particularly interesting isn’t just the benchmark performance — it’s that DeepSeek published the recipe. Training LLMs to “think longer” during inference was OpenAI’s secret sauce, and now it’s public. The stock market reaction was immediate and dramatic, but the lasting impact is structural: reasoning-capable models are no longer locked behind a proprietary API. Any team with the right hardware can run one, fine-tune it, and build on it. The question now shifts from “can you access this capability” to “what will you build with it.”