DeepSeek

What it is

DeepSeek is a Chinese AI lab whose open-weight models (DeepSeek-V3, DeepSeek-R1) rival proprietary frontier models on reasoning and coding while being freely available and dramatically cheaper to run. The models can be self-hosted, and power many BYOK setups (including Cline, Aider, and local pipelines).

Best for

• Run frontier-tier reasoning without vendor lock-in
• Power coding agents cheaply via API
• Self-host a strong model on your own GPUs

Quick start

1. Quick start: download the DeepSeek app (iOS / Android / macOS / Windows / Web) → sign up → chat with DeepSeek-V3 / R1 for free.
2. For API: visit platform.deepseek.com → create an API key (no minimum, pay-as-you-go in USDT or card).
3. Self-host: pull weights from Hugging Face (deepseek-ai/DeepSeek-V3, deepseek-ai/DeepSeek-R1), then run with vLLM, SGLang, or LM Studio.
4. Hardware: V3 is a 670B MoE — full-quality inference wants H200 / H100 nodes. R1 distill (1.5B–70B) runs on 24 GB GPUs (RTX 4090).
5. For BYOK in agents: set DeepSeek as an OpenAI-compatible endpoint at api.deepseek.com (paid) or via Ollama (local).
6. Watch the pricing — DeepSeek aggressively undercuts US labs (down to ~$0.07/M input tokens with cache hits, off-peak).

Sample input / output

Local LM Studio config:
  Model: deepseek-r1-distill-qwen-32b-q4_k_m.gguf
  ctx: 32768
  gpu: RTX 4090 (24 GB)
  temperature: 0.6
  prompt: "Solve this integral, step by step: ∫₀^∞ (1 + x²) / (eˣ − 1) dx. Show reasoning before the answer."

API config:
  POST https://api.deepseek.com/chat/completions
  {
    "model": "deepseek-reasoner",
    "messages": [{"role":"user","content":"Solve ∂u/∂t = k ∂²u/∂x² with u(0,t)=u(L,t)=0, u(x,0)=sin(πx/L)."}]
  }

Local (R1 distill, 32B q4_k_m): 38 tok/s on RTX 4090. Math proof runs 8.4s, arrives at ζ(2)/2 = π²/12 ≈ 0.822.

API (deepseek-reasoner): 9.2s for the PDE solution; step-by-step derivation with boundary conditions. Cost: $0.0028 (reasoner cache hit on the boundary term).

Comparison: same PDE on GPT-4o took 6.8s, cost $0.18. DeepSeek wins on cost-per-reasoning-step by ~50×.

Benchmarks

Pricing

Free chat; open weights; API very cheap

Underlying models

Self-hosting

Yes — DeepSeek can be self-hosted under the MIT (weights) license. This gives you full control over data and deployment.