DeepSeek R1
DeepSeek's reasoning model. Distilled 32B is excellent for 24GB GPUs. Full 671B MoE is frontier-class.
What DeepSeek R1 is for
DeepSeek R1 is a reasoning model: it spends output tokens working through a problem before answering, which makes it slower per useful word and better on problems that need the working shown. DeepSeek released it in 2025-01 with 7 sizes (1.5B, 7B, 8B, 14B, 32B, 70B, 671B) under MIT, with a 64K context window. Budget for the extra tokens the thinking phase consumes. The catalogue records its strength as reasoning, math, step-by-step problem solving.
Sizes and memory
The 7B needs about 5 GB of weights at Q4_K_M and the 671B about 384 GB, a spread of 379 GB across the family. The catalogue's default pick is 32B. Recommended VRAM adds roughly half again on top of the weights for the KV cache and the runtime.
| Size | Weights at Q4_K_M | Recommended VRAM | Smallest card in our guide that holds it |
|---|---|---|---|
| 7B | 5 GB | 8 GB | NVIDIA Jetson Orin Nano 8GB |
| 14B | 9 GB | 14 GB | NVIDIA RTX 3060 12GB |
| 32B | 20 GB | 30 GB | NVIDIA RTX 3090 |
| 70B | 42 GB | 63 GB | Apple M3 Ultra |
| 671B | 384 GB | 576 GB | None in this guide |
Hardware that runs DeepSeek R1
Every one of the 19 cards with dedicated memory in our hardware guide holds at least one size of DeepSeek R1 at Q4_K_M. The cheapest that does is the NVIDIA Jetson Orin Nano 8GB at about $199, running the 7B.
The largest size any of them holds is the 70B at 42 GB, on the Apple M3 Ultra (192 GB at 800 GB/s). Bandwidth, not capacity, sets the speed once a size fits, so a card further down the list with faster memory will generate more tokens per second on the same weights than a larger, slower one.
What people use it for
The catalogue lists 4 recorded uses for DeepSeek R1: reasoning, math, code, research. It is tagged reasoning, thinking, math, code, chain-of-thought.
Where DeepSeek R1 sits in the DeepSeek family
Our catalogue holds 4 DeepSeek entries. Ordered by release date, DeepSeek R1 (2025-01) is the 4th of them, and the most recent. DeepSeek Coder carries the same recorded memory footprint, so choosing between it and this one is a question of behaviour rather than of hardware: DeepSeek R1 is the one recorded here for reasoning, math, step-by-step problem solving.
The MIT licence puts no usage ceiling on it, which matters if the thing you are building has users. That is the practical difference between this and a vendor community licence.
Context window: 64K
64K tokens is roughly 48,000 words at the usual ratio, which is a small repository or a whole book. Two cautions come with a window this size: the KV cache at full length can rival the weights for memory, and models rarely use the far end of their advertised context as well as they use the near end. Treat it as headroom, not as a promise.
How to run DeepSeek R1 locally
Our catalogue does not record registry tags, so look the current tag up in the Ollama library or on Hugging Face before pasting these. The commands below are the shape of the workflow, not a copy-and-paste recipe.
Option 1: Ollama (simplest)
# find the tag for DeepSeek R1 at ollama.com/library
ollama run <tag>Option 2: Mullama (production)
mullama pull <tag>
mullama run <tag>Option 3: llama.cpp (CLI)
# download a GGUF from Hugging Face, searching for "DeepSeek R1 GGUF"
./llama-cli -m model.Q4_K_M.gguf -p "Hello, AI!"Option 4: Python with Mullama or llama-cpp-python
from mullama import Model, Context
model = Model.load("model.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))Sources
Model record from src/data/models.json (2026-06-29); card capacities from src/data/gpus.json (2026-06-29). Upstream: Curated from ollama.com/library + community benchmarks + paperswithcode + huggingface.