DeepSeek R1

deepseek-r1

DeepSeek R1 is the reasoning-focused DeepSeek model, widely referenced for open reasoning traces and strong math, logic, and coding performance. It shares the large MoE profile of the V3 generation but is trained and presented around deliberate reasoning rather than general chat alone. It is best described for tasks where the answer requires decomposition, verification, or step-by-step problem solving.

Context Window

128K tokens

Maximum Output

32.8K tokens

Release Date

Jan 20, 2025

Modalities

DeepSeek R1 Pricing

Input Price	Output Price	Cache Read
$0.5714/M	$2.2857/M	$0.2286/M

DeepSeek R1 API Capabilities

Reasoning

Supported

Tool calling

Supported

Temperature parameter

Supported

Attachments

Not supported

Knowledge Base

2024-07-01

Endpoint Protocols

Completions APIMessages API

DeepSeek R1 Model Highlights

DeepSeek R1 is an open reasoning model designed for extended problem solving across mathematics, code and general reasoning.

RL-Trained Reasoning

Large-scale reinforcement learning develops extended reasoning patterns for working through difficult problems.

Reflection and Verification

Its reasoning process can revisit intermediate steps and verify candidate approaches before settling on an answer.

Math and Code Reasoning

Evaluations demonstrate targeted capability in mathematical problem solving, competitive programming and software tasks.

DeepSeek R1 Use Cases

DeepSeek R1 is suited to mathematical problem solving, algorithmic programming and research tasks that require explicit multi-step reasoning.

Advanced Mathematics

Breaks down difficult mathematical problems, evaluates alternative approaches and produces a reasoned solution.

Algorithmic Programming

Designs algorithms, reasons about correctness and complexity, and generates implementations for programming problems.

Reasoned Research Analysis

Compares hypotheses, examines supplied evidence and develops a structured conclusion for technical research questions.

How to Use DeepSeek R1 via the TokenHub API

Create API key

import OpenAI from "openai"

const client = new OpenAI({
  apiKey: process.env.TOKENHUB_API_KEY,
  baseURL: "https://us-api.tokenhub.com/v1",
})

const result = await client.chat.completions.create({})
console.log(result.choices[0]?.message?.content)

DeepSeek R1 Benchmarks

DeepSeek R1 (Jan '25)

Index score
Artificial Analysis Intelligence IndexArtificial Analysis Intelligence Index summarizes broad model capability across evaluation families.	Artificial Analysis broad capability aggregate	12.6
Artificial Analysis Coding IndexArtificial Analysis Coding Index summarizes model performance on coding and software engineering tasks.	Artificial Analysis software task aggregate	15.9
Artificial Analysis Math IndexArtificial Analysis Math Index summarizes mathematical reasoning and competition-style problem solving.	Artificial Analysis math reasoning aggregate	68
Knowledge & Reasoning
MMLU-ProMMLU-Pro evaluates broad academic and professional knowledge with more challenging questions than the original MMLU.	Advanced multi-task knowledge	84.4%
GPQAGPQA is a graduate-level science benchmark designed to test expert reasoning.	Advanced science problem solving	70.8%
HLEHLE is a broad, challenging benchmark for expert-level knowledge and reasoning across many domains.	Broad expert-level exam set	9.3%
Coding & Engineering
LiveCodeBenchLiveCodeBench evaluates coding ability on recent programming problems designed to reduce benchmark contamination.	Live coding problems	61.7%
SciCodeSciCode evaluates Python programming for scientific computing, including algorithmic implementation from scientific context.	Scientific coding challenges	35.7%
Terminal-Bench HardTerminal-Bench Hard evaluates whether an AI agent can complete difficult tasks in a terminal environment.	Hard terminal task execution	6.1%
Math
MATH-500MATH-500 evaluates mathematical problem solving across a curated set of challenging questions.	Advanced math problem solving	96.6%
AIMEAIME measures mathematical reasoning on competition-style problems.	Competition math problems	68.3%
AIME 2025AIME 2025 measures performance on challenging competition math problems.	Competition math problems	68%
Instruction Following & Agent Tasks
IFBenchIFBench measures how reliably a model follows detailed instructions and constraints.	Prompt constraint adherence	39.0%
AA-LCRAA-LCR measures reasoning quality over extended context windows.	Long-context reasoning	52.3%
τ²-Benchτ²-Bench evaluates conversational agents in realistic workflows with tool and state constraints.	Agent workflow tasks	11.4%

Metrics sourced from Artificial Analysis

DeepSeek R1 FAQ

DeepSeek R1: capabilities, use cases, limits, and TokenHub guidance.

What is DeepSeek R1 built for?+

DeepSeek R1 is a DeepSeek model for open-weight, reasoning-intensive problem solving.

When is DeepSeek R1 most useful?+

Best for mathematical reasoning, code reasoning and scientific reasoning, especially when deep reasoning is the priority.

What distinguishes DeepSeek R1?+

Key strength: reasoning-focused post-training with openly released weights.

What should I watch for?+

Deep reasoning can increase response time and token use. For the latest capabilities matter, consider DeepSeek V4 Pro.

How do I check availability?+

Use TokenHub's exact ID; hosted behavior may differ from self-hosting.

Ready to use DeepSeek R1?

Use one API key to access DeepSeek R1 and more AI models through TokenHub.

Create API key

DeepSeek R1

DeepSeek R1 Pricing

DeepSeek R1 API Capabilities

DeepSeek R1 Model Highlights

RL-Trained Reasoning

Reflection and Verification

Math and Code Reasoning

DeepSeek R1 Use Cases

Advanced Mathematics

Algorithmic Programming

Reasoned Research Analysis

How to Use DeepSeek R1 via the TokenHub API

DeepSeek R1 Benchmarks

Models Similar to DeepSeek R1

Media and Discussions

X (Twitter)

Reddit

YouTube

DeepSeek R1 FAQ

Ready to use DeepSeek R1?