Reasoning models, such as DeepSeek-R1, systematically process information to derive logical conclusions and solve complex problems step by step. These models are trained to output a long chain of thoughts via reinforcement learning and have the ability to self-correct. This capability transforms AI to move beyond simple response generation, allowing for multi-step reasoning and verification, resulting in more accurate and reflective outputs.

​
Model information

DeepSeek R1 is a 671 billion parameter open-source reasoning model. It utilizes tokens specifically for reflecting on how to respond to the user prompt, shown within the <think>tags in an output. Due to this difference of structure from typical model requests, consider the following alterations.

​
System prompt

  • Remove instructional system prompts and minimize additional prompting beyond the user query. Excessive instructions can restrict the model’s reasoning scope and reduce output quality.

​
User prompt

  • Avoid additional chain of thought prompting and explicating how to respond since the model already processes queries in its thinking approach. It’s best to use zero-shot or single-instruction plain language prompts for complex tasks. This approach allows the model’s internal reasoning capabilities to shine.
  • Experiment with the structure of these sections: goal, return format, warnings, and context dump.
  • In the rare case the <think> tags are bypassed, you can enforce using them by telling the model to start with <think> tags.
  • To generate concise thinking, the idea of chain-of-draft can be used, that is, such as adding “only keep a minimum draft for each thinking step, with 5 words at most” as a constraint in your prompt.
Depiction of reasoning versus non-reasoning model responses.

​
Parameters

  • For general reasoning (non-math reasoning), the suggested temperature = 0.6 and top-p = 0.95. If you prefer more factual response, temperature can be set as smaller one, such as 0.5.
  • For math-style reasoning, the suggested temperature is 0.7 and top-p = 1.0.

​
Example request

import OpenAI from "openai";
const client = new OpenAI({
  baseURL: "your-sambanova-base-url",
  apiKey: "your-sambanova-api-key",
});
const chatCompletion = await client.chat.completions.create({
  messages: [
    { role: "system", content: "Answer the question in a couple sentences." },
    { role: "user", content: "I want to understand the factors I should consider when buying a house. For context, I live in California and am 30 years old." },
  ],
  model: "DeepSeek-R1",
});
console.log(chatCompletion.choices[0].message.content);

​
Use cases

​
Report generator

DeepSeek-R1 is particularly adept at processing unstructured information, making it ideal for analyzing complex documents like legal contracts, financial statements, or scientific papers. Reasoning models can facilitate pattern recognition by analyzing multiple facets of the information, prior to distilling it into a comprehensive summary. Example prompt Develop a comprehensive report on the state of autonomous vehicles. Present this report with organized sections and a breif summarization. Be careful to cite the achievements with the proper entity that made that achievement or contribution. For context: I am knowledgable in this field and have a technical understanding of autonmous vehicle systems. I've been working most of my career in artificial intelligence but have not yet joined a company with the sole focus of autonomous vehicles. I am considering making the career change and wanted to understand the current ecosystem before I go through the job search process.

​
Planner for workflows and agents

Reasoning models excel at tackling ambiguous and complex tasks. They can break down intricate problems, strategize solutions, and make decisions based on large volumes of ambiguous information. Use DeepSeek-R1 as a strategic planner for complex, multi-step problems. It can break down tasks, develop detailed solutions, and even orchestrate other AI models for specific subtasks within an agentic system. To visualize the vast capabilities of thoughtful planning in a powerful workflow, the demo app implements DeepSeek-R1 as a planner. It orchestrates agents with various workflows, such as Deep Research, Financial Analysis, and Sales Leads. The app is open-source allowing developers quick experimentation and easy production of their own agent workflows within the system.

​
Coding and mathematical guru

These models are effective at reviewing and improving code. They can detect minor changes in a codebase that might be missed by human reviewers. In solving math problems, reasoning is helpful to breakdown tasks into many steps and verify its work throughout solving.

​
Best Practices

Latency and cost

​
FAQs