DeepSeek DeepSeek Released 2025-01 reasoningthinkingmath

DeepSeek R1

DeepSeek's reasoning model. Distilled 32B is excellent for 24GB GPUs. Full 671B MoE is frontier-class.

Best for reasoning, math, step-by-step problem solving
Sizes 1.5B · 7B · 8B · 14B · 32B · 70B · 671B
Context 64K
License MIT
Min VRAM (default size, Q4) 1 GB
Rec VRAM 8 GB

What is DeepSeek R1?

DeepSeek R1 is DeepSeek's open-weights language model. Released in 2025-01, it ships in 7 sizes (1.5B, 7B, 8B, 14B, 32B, 70B, 671B) and is licensed MIT. The most popular use case is reasoning, math, step-by-step problem solving.

VRAM and hardware

The smallest 1.5B size needs at least 1 GB VRAM; the largest 671B needs around 5 GB.

Size Min VRAM (Q4_K_M) Recommended VRAM Tokens/sec on 3090
1.5B 1 GB 2 GB ~34 tok/s
7B 5 GB 8 GB ~25 tok/s
8B 1 GB 2 GB ~34 tok/s
14B 9 GB 14 GB ~22 tok/s
32B 20 GB 30 GB ~17 tok/s
70B 42 GB 63 GB ~14 tok/s
671B 384 GB 576 GB ~9 tok/s

How to run DeepSeek R1 locally

Option 1: Ollama (simplest)

ollama pull deepseek-r1
ollama run deepseek-r1

Option 2: Mullama (production)

mullama pull deepseek-r1
mullama run deepseek-r1

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "deepseek-r1 gguf")
./llama-cli -m deepseek-r1.Q4_K_M.gguf -p "Hello, AI!"

Option 4: Python with Mullama or llama-cpp-python

from mullama import Model, Context
model = Model.load("deepseek-r1.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))

What you can build with DeepSeek R1

  • reasoning
  • math
  • code
  • research

Hardware it fits on

  • Apple Silicon (8GB): 1.5B in MLX
  • Apple Silicon (16GB): 7B in MLX
  • Apple Silicon (32GB): 8B in MLX or GGUF
  • Apple Silicon (64GB+): 671B in MLX
  • RTX 3090 (24GB): 8B at Q4_K_M
  • RTX 4090 (24GB): same as 3090, ~30% faster
  • RTX 5090 (32GB): 8B at Q8_0
  • 2× RTX 6000 Ada (96GB total): 671B at Q4_K_M

Related models in the DeepSeek family

  • DeepSeek V3 — 671B (37B active MoE) · frontier-class open-weights LLM (needs data-center GPUs)
  • DeepSeek Coder — 1.3B, 6.7B, 33B · small-footprint code completion
  • DeepSeek Coder V2 — 16B, 236B · strong 16B code model with MoE efficiency

Other models you might consider

  • Llama 3.1 (Meta) — 8B, general-purpose chat, agentic workflows, code
  • Llama 3.2 (Meta) — 1B, low-end hardware, mobile, edge deployment
  • Llama 3.2 Vision (Meta) — 11B, vision + chat on desktop GPUs
  • Llama 3.3 (Meta) — 70B, best-in-class 70B for desktops
  • Llama 4 (Meta) — Scout 17B (109B MoE), vision + long context (10M tokens)
  • Qwen 2.5 (Alibaba) — 0.5B, strongest 7B-32B on consumer hardware