This guide covers running DeepSeek R1 on Runpod, from the full 671B reasoning model down to the smaller distilled variants. DeepSeek R1 is an open-weight reasoning model from DeepSeek AI that works through problems step by step before answering, with performance that holds up against closed reasoning models on math, coding, and logic benchmarks.
This guide walks through what DeepSeek R1 actually is, which variant fits your GPU budget, and how to get one running on Runpod, whether that's a one-click Serverless endpoint or a full Pod deployment. By the end, you'll have a working DeepSeek R1 endpoint you can send requests to.
What is DeepSeek R1?
DeepSeek R1 is a reasoning model released by DeepSeek AI in January 2025. Unlike a standard chat model, R1 generates an internal chain of reasoning before producing its final answer, which is what drives its results on math, coding, and multi-step logic tasks. It's released under the MIT license, so the weights are open for commercial use, modification, and distillation.
The full model uses a Mixture-of-Experts architecture with 671B total parameters, of which about 37B are active per token, the same underlying architecture as DeepSeek V3. Alongside the full model, DeepSeek released a set of distilled variants: smaller, dense models fine-tuned on reasoning traces generated by R1 itself, built on top of Qwen2.5 and Llama 3 base models.
- DeepSeek-R1 (671B): The full reasoning model. Highest capability, requires a multi-GPU setup.
- DeepSeek-R1-Distill-Qwen (1.5B / 7B / 14B / 32B): Dense models distilled from R1 onto Qwen2.5 bases. Practical for single-GPU deployment.
- DeepSeek-R1-Distill-Llama (8B / 70B): Dense models distilled from R1 onto Llama 3 bases. The 70B variant trades some of the smaller models' speed for stronger reasoning.
The distilled models aren't the full model compressed, they're separate, smaller models trained to imitate R1's reasoning behavior. That tradeoff is what makes them worth running: most of R1's step-by-step reasoning quality at a fraction of the hardware requirement.
Why run DeepSeek R1 on Runpod?
The full R1 model needs enough VRAM that most people will never own the hardware to run it, and even the distilled variants need a GPU most laptops don't have. Runpod removes that constraint: pick the GPU that matches the variant you want, deploy it, and pay only for the time you use it.
For the distilled variants, Runpod Serverless has ready-made endpoints you can deploy directly, with no setup:
- DeepSeek-R1-Distill-Qwen-7B
- DeepSeek-R1-Distill-Qwen-14B
- DeepSeek-R1-Distill-Qwen-1.5B
- DeepSeek-R1-Distill-Llama-8B
If you want the full 671B model or a variant not listed above, the template approach below gives you that control. The endpoints are the fastest way to get a reasoning model answering questions right now.
GPU requirements for DeepSeek R1
VRAM needs shift with quantization, so treat these as practical starting points rather than hard numbers:
- 8-16 GB VRAM (RTX 4080, RTX 4090 shared, A5000): Distill-Qwen-1.5B and 7B
- 16-24 GB VRAM (RTX 4090, A5000, L40S): Distill-Qwen-14B, Distill-Llama-8B, with room for longer context windows
- 40-48 GB VRAM (A6000, L40S, A100 40GB): Distill-Qwen-32B
- 80 GB+ VRAM (A100 80GB, H100): Distill-Llama-70B, or multiple smaller distills served concurrently
- 480 GB+ VRAM (6x A100 80GB, or 4x H200): The full 671B R1 model, 4-bit quantized
Templates for running DeepSeek R1 on Runpod
vLLM (official Runpod Serverless worker) The Runpod Hub's vLLM worker supports DeepSeek R1 and its distilled variants directly, including deepseek-ai/DeepSeek-R1-Distill-Qwen-7B and the rest of the distill family. This is the recommended path for any distilled variant, since it doesn't require managing a Pod at all.
KoboldCpp - DeepSeek-R1 - 6x80GB of VRAM A community template for running the full 671B model on a Pod. It deploys a 4-bit quantized (Q4) build of R1 through KoboldCpp, exposing an OpenAI-compatible endpoint. Requires a Pod configuration with at least 480 GB of VRAM (6x A100 80GB or 4x H200). Community templates can have bugs, review the configuration before deploying.
Step 1: Create a Runpod account
Go to runpod.io and sign up with email, GitHub, or Google. Add a payment method so you can deploy a GPU.
Step 2: Deploy DeepSeek R1
Which path you take depends on whether you want a distilled variant (fastest to deploy, least GPU-intensive) or the full model (highest capability, most hardware).
Option A: Deploy a distilled model on Serverless
- Go to the model's Runpod page (linked above) and click Deploy now, or go directly to the vLLM worker in the Runpod Hub and click Deploy.
- In the Model field, enter the Hugging Face model name, for example deepseek-ai/DeepSeek-R1-Distill-Qwen-7B.
- Click Advanced to expand the vLLM settings and set Max Model Length to match the context window you need (8192 is a reasonable default).
- Leave the remaining settings at their defaults unless you have a specific reason to change them, then click Create Endpoint.
Your endpoint will initialize while Runpod provisions the GPU and downloads the model.
Option B: Deploy the full R1 model on a Pod
- Go to the Pods section and select a GPU configuration with at least 480 GB of VRAM: 6x A100 80GB or 4x H200.
- Click Change Template, search for KoboldCpp - DeepSeek-R1 - 6x80GB of VRAM, and select it.
- Set your disk size to accommodate the quantized model files, then click Deploy.
The Pod itself deploys in under 30 seconds. Downloading and loading a model this size takes considerably longer, budget around 20 minutes for the full setup.
Step 3: Send your first request
Once your endpoint or Pod shows a running state, find your endpoint ID (Serverless) or your exposed port (Pod, KoboldCpp listens on port 5001) from the Runpod console.
For a Serverless endpoint, send a test request:
curl -X POST "https://api.runpod.ai/v2/YOUR_ENDPOINT_ID/runsync" \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"input": {"prompt": "Explain the difference between a stack and a queue."}}'
The response includes both R1's reasoning trace and its final answer. Expect the reasoning portion to be substantially longer than the answer itself, that's expected behavior for this model, not a sign anything is wrong.
Step 4: Tune reasoning behavior
DeepSeek's model documentation recommends a few settings specific to how R1 reasons:
- Temperature 0.5-0.7 (0.6 is the commonly recommended starting point). Too low, and the model can repeat or loop; too high, and reasoning quality degrades.
- Skip the system prompt. R1 was trained without one. Put any instructions directly in the user turn instead.
- For math problems, prompt the model to reason step by step and place the final answer in a boxed format, this matches how R1 was trained to structure output.
Some serving setups can separate the reasoning trace from the final answer automatically instead of returning both as raw text. Check your vLLM worker's environment variables for a reasoning-parser option if you want structured output rather than parsing <think> tags yourself.
Pro tips
Attach a network volume for repeat use. Pod storage is ephemeral by default. A network volume keeps the downloaded model persistent across sessions, so you're not re-downloading hundreds of gigabytes every time you redeploy.
Start with a distilled model, not the full 671B. Unless you specifically need the full model's capability, a 14B or 32B distill gets you most of the reasoning quality at a fraction of the cost and setup time.
Watch your context length. Reasoning traces consume tokens before the model ever gets to its answer. If you're hitting context limits, raising MAX_MODEL_LEN matters more for R1 than it does for non-reasoning models.
Stop your Pod when you're done. Runpod bills by the minute. A 6-GPU Pod running the full model idle overnight adds up fast.
Wrapping up
DeepSeek R1 made open-weight reasoning models practical, and Runpod is the fastest path to running one, whether that's a distilled model on a Serverless endpoint in a few minutes or the full 671B model on a dedicated Pod. Start with a distilled variant if you're not sure which one you need, you can always move up to a larger model or the full R1 once you know your workload.
Ready to try it? Deploy a distilled model for the fastest start, or deploy the full R1 template for maximum capability.
FAQ
What's the difference between DeepSeek R1 and the distilled models?
R1 is the full 671B Mixture-of-Experts reasoning model. The distilled variants are smaller, dense models (based on Qwen2.5 or Llama 3) fine-tuned on reasoning traces generated by R1, not compressed versions of R1 itself. They retain much of R1's step-by-step reasoning quality while running on a fraction of the hardware.
What GPU do I need to run DeepSeek R1 on Runpod?
It depends on the variant. The smallest distills run on 8-16 GB VRAM, the 32B distill needs around 40-48 GB, the 70B distill needs 80 GB or more, and the full 671B model needs roughly 480 GB of VRAM at 4-bit quantization. Runpod lets you rent exactly the GPU configuration each variant needs, without buying the hardware.
Can I run DeepSeek R1 locally instead?
The smaller distills, 1.5B through 14B, can run on a single consumer GPU with enough VRAM. The 70B distill and the full 671B model need hardware most individuals don't own. Running on Runpod avoids the upfront hardware cost and the setup work of managing drivers, quantization, and dependencies yourself.
Is DeepSeek R1 free to use commercially?
Yes. DeepSeek R1 and its distilled variants are released under the MIT license, which permits commercial use, modification, and further distillation.
Can I use DeepSeek R1 as an API instead of a UI?
Yes. Runpod Serverless deploys DeepSeek R1 (or any of its distilled variants) as a REST API that autoscales automatically. Send requests to your endpoint the same way you would to any hosted LLM API, and pay only for the compute time each request uses.
