“DeepSeek this, DeepSeek that” — the buzz around DeepSeek’s large language models is electrifying! With groundbreaking advancements in AI, DeepSeek is revolutionizing industries across the globe. We’re excited to share that DeepSeek’s powerful models are now accessible on FriendliAI’s platforms. DeepSeek R1 671B is ready for use on Serverless Endpoints. Even better, you can deploy optimized, distilled, and fine-tuned DeepSeek models on Dedicated Endpoints!

But what makes DeepSeek so special, and how does FriendliAI supercharge models? Let’s dive in.

What is even DeepSeek?

DeepSeek models, developed by the Chinese AI company DeepSeek, have gained significant attention due to their competitive performance and open-source availability. Key models include:

• DeepSeek V3 (December 2024): A 671B parameter language model using Mixture-of-Experts (MoE) architecture and Multi-head Latent Attention (MLA). It features auxiliary-loss-free strategy and Multi-Token Prediction (MTP) for improved performance. Post-trained using Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), it competes with closed-source models like Claude 3.5 Sonnet on certain benchmarks.
• DeepSeek R1 (January 2025): Built upon V3, improving reasoning and problem-solving through large-scale RL. While R1 Zero, trained with RL without SFT, showed impressive reasoning, it encountered issues like endless repetition, poor readability, and language mixing. R1 incorporated cold-start data before training to address these issues. R1 outperforms OpenAI’s o1 model on some benchmarks, and its distilled version, DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI’s o1-mini on several benchmarks.

Open-source models like DeepSeek make cutting-edge AI accessible, allowing users to build on existing frameworks without proprietary constraints. This flexibility helps individuals, developers, and organizations create tailored AI solutions for diverse challenges.

Why Choose FriendliAI?

While DeepSeek’s AI models offer impressive capabilities, the company’s app has raised security and privacy concerns, leading to bans in several countries. This is where FriendliAI comes into play.

FriendliAI stands out as the premier choice for utilizing DeepSeek models, offering enhanced security by keeping your data safe. With pioneering inference acceleration technology, FriendliAI boosts processing speeds, ensuring efficient AI operations.

Recognized globally as the first startup to be integrated as a deployment option on HuggingFace, FriendliAI combines security, speed, affordability, and customization, making it easy to scale and integrate DeepSeek models into your applications, offering flexible options for diverse customer needs.

Leveraging DeepSeek models is easier than ever with Friendli Suite. Choose the option that suits your needs:

Use Serverless Endpoints

Start instantly by sending API requests or visit Friendli Suite Serverless Endpoint Playground:

Figure 1: Running DeepSeek R1 on Friendli Suite Serverless Endpoint Playground.

For example, you can use OpenAI Python SDK to interact with FriendliAI, along with many other popular SDKs supported. Click here to see more.

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.friendli.ai/serverless/v1",
    api_key=os.environ.get("FRIENDLI_TOKEN")
)

completion = client.chat.completions.create(
    model="deepseek-r1",
    messages=[{"role": "user", "content": "How to make a taco?"}]
)

print(completion.choices[0].message)

Learn more about how to use Serverless Endpoints here.

Deploy on Dedicated Endpoints

Quickly deploy optimized distilled models and custom fine-tuned models from HuggingFace with ease. To deploy models on Dedicated Endpoints, simply click the “Deploy” button on HuggingFace.

Figure 2: How to deploy with Friendli Inference from HuggingFace.

Here are the links to some of the models, provided for your convenience:

• DeepSeek-R1-Distill-Llama-70B
• DeepSeek-R1-Distill-Llama-8B
• DeepSeek-R1-Distill-Qwen-32B
• DeepSeek-R1-Distill-Qwen-14B
• DeepSeek-R1-Distill-Qwen-7B
• DeepSeek-R1-Distill-Qwen-1.5B

You will then be directed to our Deploy Model page; click the “Deploy now” button. This action will prompt you to choose your deployment method: opt for a quick setup with “One-click” or customize your deployment with “Configure myself.” This simple and intuitive process lets you launch your models efficiently with minimal effort. That’s it! You are all set to go. Happy deploying!

Figure 3: Deploying model on Friendli Dedicated Endpoint.

Get Started Today!

Don’t miss out on the opportunity to harness the power of DeepSeek models. Industry leaders worldwide have already chosen FriendliAI — now it’s your turn. Click here to accelerate your business with generative AI and start building today.