DeepSeek

DeepSeek provides state-of-the-art open foundation models, including DeepSeek-V3 (Mixture-of-Experts) and DeepSeek-R1 for chain-of-thought mathematical reasoning and agentic workflows.

When to Use

• High-Performance Open-Weights Reasoning & Coding: DeepSeek-R1 and DeepSeek-V3 for mathematical reasoning, agentic logic, and code generation.
• Self-Hosted & Private LLM Deployments: Running enterprise reasoning models locally using vLLM, SGLang, or Ollama.
• Cost-Efficient API Scaling: Accessing state-of-the-art reasoning at a fraction of proprietary API costs.
• Chain-of-Thought (CoT) Verification: Parsing detailed <think> reasoning traces for auditability and verification.

Quick Start

from openai import OpenAI

# DeepSeek exposes an OpenAI-compatible API endpoint
client = OpenAI(
    api_key="your-deepseek-api-key",
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-reasoner", # DeepSeek-R1
    messages=[
        {"role": "user", "content": "Solve: How many r's are in strawberry? Think step by step."}
    ]
)

# Reasoning output is provided in reasoning_content
print("Thinking Process:\n", response.choices[0].message.reasoning_content)
print("Final Answer:\n", response.choices[0].message.content)

Core Concepts

Consuming DeepSeek API with OpenAI SDK Compatibility

Querying DeepSeek models with reasoning token handling:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-reasoner", # DeepSeek-R1 reasoning model
    messages=[
        {"role": "system", "content": "You are a quantitative systems engineer."},
        {"role": "user", "content": "Design an optimal concurrent lock-free queue algorithm in C++."}
    ],
    max_tokens=4096,
    temperature=0.6,
)

# DeepSeek-R1 returns reasoning content separately
reasoning = getattr(response.choices[0].message, 'reasoning_content', None)
final_answer = response.choices[0].message.content

if reasoning:
    print(f"=== Thinking Process ({len(reasoning)} chars) ===")
    print(reasoning[:500] + "...")

print("\n=== Final Response ===")
print(final_answer)

High-Throughput Self-Hosting with vLLM

Deploying DeepSeek-V3 / R1 on multi-GPU nodes with PagedAttention:

# Launch vLLM server with tensor parallelism across 8 H100 GPUs
vllm serve deepseek-ai/DeepSeek-R1 \
  --tensor-parallel-size 8 \
  --max-model-len 32768 \
  --trust-remote-code \
  --gpu-memory-utilization 0.95 \
  --port 8000

Streaming Reasoning Tokens

Streaming live thinking tokens to the client interface:

response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=[{"role": "user", "content": "Solve this riddle: ..."}],
    stream=True
)

for chunk in response:
    delta = chunk.choices[0].delta
    if hasattr(delta, 'reasoning_content') and delta.reasoning_content:
        print(f"[Thinking]: {delta.reasoning_content}", end="", flush=True)
    elif delta.content:
        print(delta.content, end="", flush=True)

Common Patterns

DeepSeek-R1 Reasoning Extraction with Fallback

Problem: Handling reasoning steps separately from final response text in client applications.

Solution: Access reasoning_content with safe attribute fallback:

msg = response.choices[0].message
thought_process = getattr(msg, 'reasoning_content', None)
final_text = msg.content

if thought_process:
    print(f"Step-by-step logic: {thought_process}")
print(f"Output: {final_text}")

Best Practices

Do:

• Separate reasoning output (reasoning_content) from final output (content) when rendering responses to users.
• Use temperature=0.6 (recommended default for DeepSeek-R1) to balance logical rigor and exploration.
• Use FP8 or AWQ 4-bit quantizations when self-hosting on hardware with limited VRAM.
• Implement retries with exponential backoff on API endpoints during peak network congestion.

Don't:

• Strip <think> tags prematurely if debugging algorithmic reasoning failures.
• Provide overly verbose system prompts for DeepSeek-R1; it is trained to reason autonomously.
• Use temperature=0 with DeepSeek reasoning models; it may cause repetitive reasoning loops.

Troubleshooting

References

• DeepSeek GitHub