The original DeepSeek R1 has 671 billion total parameters. It is a Mixture-of-Experts model, so about 37 billion parameters are activated for each token. Those numbers describe the model’s parameter count—not the size of a download or the amount of memory your computer needs to run it. DeepSeek’s R1 repository says R1 is based on DeepSeek-V3-Base; DeepSeek’s V3 technical report gives the 671B total and 37B-per-token figures. (DeepSeek-R1 repository)
What does “671 billion parameters” mean?
A parameter is a learned value in a model. DeepSeek R1’s 671B figure counts the model’s parameters overall. The 37B active figure describes how many are used for an individual token. Because R1 is a Mixture-of-Experts model, these figures are different—but the active count does not mean the remaining weights can simply be left out of the model. (DeepSeek-V3 repository)
DeepSeek also releases smaller models distilled from R1, including versions labeled 1.5B, 7B, 8B, 14B, 32B, and 70B. Those are separate, smaller models—not the original 671B R1—so their size or hardware guidance should not be mistaken for the full model’s. (DeepSeek-R1 README)
Is the model 671 GB or 400 GB to download?
Not necessarily. Parameter count, weight-file size, and runtime memory are related but not interchangeable. The actual weight file depends on the format and precision used to store the parameters; running the model also involves memory for inference and can require additional space for context and other runtime data.
As a rough illustration—not an official download-size specification—671 billion values stored at 16 bits each would amount to about 1.34 TB of weight data before other overhead. At 4 bits each, the simple arithmetic is about 336 GB, also before overhead. Actual model artifacts can differ, so these calculations should not be treated as exact file sizes or memory requirements.
Quickchat AI reports approximately 400 GB for full R1 and 131 GB for a quantized version, but does not identify the precise artifact, format, or whether those figures mean disk space or runtime memory. Treat them as Quickchat AI’s estimates, not confirmed download sizes. (Quickchat AI)
How much memory or hardware does it need?
There is no single memory figure in the information cited here that should be treated as a universal minimum. Requirements vary with the model artifact and precision, serving software, context length, and runtime configuration. A model’s 37B active parameters alone do not tell you how much memory the complete model needs.
For a practical choice, first distinguish the original 671B model from a distilled R1 model. If you intend to run the original locally, check the requirements for the exact weight artifact and serving setup you plan to use; do not rely on a VRAM figure for a smaller distill. If you do not need to operate the weights yourself, check whether a hosted API offers the specific model you require.
Can I use a production API instead of hosting R1?
An API can let an application call a hosted model without your team running its weights, but the model behind the API matters: access to a reasoning model is not automatically access to the original DeepSeek R1.
DeepSeek’s API documentation lists deepseek-flash and deepseek-v4-pro, with deepseek-flash served by DeepSeek-V4.1-Flash-1017. DeepSeek’s API model listing and pricing page are places to check for current model IDs and terms. (DeepSeek API documentation)
TokenForge’s support for the original R1 or an R1 distill is unconfirmed here. I could not verify an exact TokenForge model ID or the capabilities of any such offering. Check TokenForge’s current API catalog before choosing it; DeepSeek’s own model listing does not establish what TokenForge serves.
Bottom line: DeepSeek R1 is a 671B-parameter model, with 37B parameters activated per token. Neither number is a download-size or hardware requirement. For local deployment, evaluate the exact full-model artifact and runtime; for API use, verify that the provider offers the model you actually want—not simply an R1 distill or another reasoning model.
Sources
DeepSeek-R1 official repository
DeepSeek-V3 official repository and technical report
We tested DeepSeek. Here's what you need to know