Which Model Size to Choose
The largest one that fits entirely in your VRAM is the one you should choose. Running entirely in VRAM is fastest; it can still run if it doesn't fit, but part of it gets pushed into memory and computed on the CPU, which is much slower.
First try it with your VRAM:
$ ollama psNAME SIZE PROCESSORdeepseek-r1:8b 5.7 GB 100% GPUdeepseek-r1:8b Can be embedded entirely into 8 GB of VRAM, runs the fastest.
What You'll Be Able to Do After This Lesson
- State the file sizes of each DeepSeek-R1 size
- Pick a size based on your own VRAM
- Use
ollama psto confirm your choice is correct
How Big Is Each Size
Sizes of deepseek-r1 in the Ollama model library (ollama run deepseek-r1:<尺寸>):
| Size | File Size | Context Length |
|---|---|---|
| 1.5b | 1.1 GB | 128K |
| 7b | 4.7 GB | 128K |
| 8b | 5.2 GB | 128K |
| 14b | 9.0 GB | 128K |
| 32b | 20 GB | 128K |
| 70b | 43 GB | 128K |
| 671b | 404 GB | 160K |
If you run ollama run deepseek-r1 without specifying a size, you get latest, whose file size is the same as 8b: 5.2 GB.
The "b" stands for billions of parameters. Generally, the larger the size, the better the answers — and the more VRAM it consumes.
- I know how large the file is for the size I want to try
How to Tell Whether It Fits
Once loaded, the model uses slightly more VRAM than the file size. In our test of the 8b: the file is 5.2 GB, and ollama ps shows 5.6 GB. So as a rough rule, VRAM needs to be noticeably larger than the file. The small tool above estimates based on this.
The real answer is in the PROCESSOR column of ollama ps:
| Display | Meaning |
|---|---|
100% GPU | Entirely on the GPU, fastest |
48%/52% CPU/GPU Like this | Partly in memory, partly on the GPU, slower |
100% CPU | Entirely in memory, computed on the CPU, slowest |
If it's not 100% GPU, try one size smaller.
Look it upMy GPU is…
Click one to see.
- I've chosen a size and checked it with
ollama ps
Checkpoint
- I know the file sizes of each size
- I know VRAM needs to be noticeably larger than the file
- I can read the
PROCESSORcolumn to tell whether the model is entirely on the GPU
Common Misconceptions
Thinking bigger is always better. A large model that doesn't fit in VRAM runs much slower. Start with the largest size that fits.
You might think the R1 on this page is DeepSeek's latest model. The latest ones are the V4 / V4.1 series, but they won't fit on a home computer, and Ollama only offers the cloud version — seeCan DeepSeek V4 run locally. What can still run locally are the R1 sizes listed above.
Thinking latest is the largest version. For this model, latest has the same file size as 8b: 5.2 GB.