DeepSeek V3

What is DeepSeek V3?

DeepSeek V3 is DeepSeek's MoE open-weights language model. Released in 2024-12, it ships in 1 size (671B (37B active MoE)) and is licensed DeepSeek License. The most popular use case is frontier-class open-weights LLM (needs data-center GPUs).

VRAM and hardware

The 671B (37B active MoE) size fits in 256 GB of VRAM.

Size	Min VRAM (Q4_K_M)	Recommended VRAM	Tokens/sec on 3090
671B (37B active MoE)	256 GB	384 GB	~10 tok/s

How to run DeepSeek V3 locally

Option 1: Ollama (simplest)

ollama pull deepseek-v3
ollama run deepseek-v3

Option 2: Mullama (production)

mullama pull deepseek-v3
mullama run deepseek-v3

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "deepseek-v3 gguf")
./llama-cli -m deepseek-v3.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-v3.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 V3

• research
• frontier tasks
• offline frontier inference

Hardware it fits on

• Apple Silicon (8GB): 671B (37B active MoE) in MLX
• Apple Silicon (16GB): 671B (37B active MoE) in MLX
• Apple Silicon (32GB): 671B (37B active MoE) in MLX or GGUF
• Apple Silicon (64GB+): 671B (37B active MoE) in MLX
• RTX 3090 (24GB): 671B (37B active MoE) at Q4_K_M
• RTX 4090 (24GB): same as 3090, ~30% faster
• RTX 5090 (32GB): 671B (37B active MoE) at Q8_0
• 2× RTX 6000 Ada (96GB total): 671B (37B active MoE) at Q4_K_M

Related models in the DeepSeek family

• DeepSeek R1 — 1.5B, 7B, 8B, 14B, 32B, 70B, 671B · reasoning, math, step-by-step problem solving
• 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