DeepSeek Open Weights (MIT License)Status: active

DeepSeek-R1

Open-weights frontier reasoning model trained via large-scale reinforcement learning without supervised cold start.

Released: Jan 20, 2025
Last Verified: Aug 10, 2026
Context Window

64k

Tokens

Max Output

8.192k

Output limit

Architecture

MoE (Mixture of Experts) Reasoning

Model family

Parameters

671B (37B active)

Total / Active

Input Token Cost

$0.55

Per 1M tokens

Output Token Cost

$2.19

Per 1M tokens

Model Overview

DeepSeek-R1 is an open-weights (MIT license) 671B parameter Mixture-of-Experts (MoE) reasoning model. Trained with large-scale RL, it exhibits transparent chain-of-thought reasoning enclosed in `<think>` tags, matching OpenAI o1 on competition math (AIME 79.8%) and coding benchmarks at a fraction of inference cost. Can be run locally or via API.

Developer Implementation Notes

Transparent `<think>` tags allow inspecting internal reasoning. Full model requires ~400GB VRAM (FP8) or 8x 80GB GPUs; distilled versions (1.5B, 7B, 8B, 14B, 32B, 70B) run on consumer GPUs and laptops via Ollama.

Key Strengths

  • Open-weights under permissive MIT license
  • Uncensored and transparent chain-of-thought token visibility
  • Extremely cheap API pricing ($0.55 / $2.19 per MTok)
  • Distilled models available from 1.5B to 70B for consumer hardware

Limitations & Boundaries

  • Full 671B model requires massive multi-GPU hardware for local hosting
  • 64k context window is smaller than Claude/Gemini
  • Can occasionally switch languages mid-thought in complex loops

Capabilities & Modalities

Modalities:text
Reasoning Chains:Supported
Tool / Function Calling:Supported
Structured Outputs:Supported
Fine-Tuning:Supported
Local Deployment:Yes
Min VRAM:40 GB

Best Production Use Cases

  • Private on-premise deep reasoning & math evaluation
  • Self-hosted coding and security audit pipelines
  • Synthetic training data generation for smaller models

Verified Benchmark Results

Standardized evaluations with methodology notes and authoritative citation links.

View all industry benchmarks →
Reasoning & Math Official
AIME (2024/2025)

79.8%

Pass@1 with native <think> chain-of-thought tokens

Coding Official
SWE-bench Verified

49.2%

Pass@1 with standard SWE-bench framework

Coding Official
LiveCodeBench

65.9%

Pass@1 0-shot with reasoning tokens

Reasoning & Math Official
GPQA Diamond

71.5%

Zero-shot with reasoning tokens

General Knowledge & Frontier Official
MMLU-Pro

84.0%

5-shot with full reasoning chain

Pricing & Inference Cost Calculator

Token Cost Estimator – DeepSeek-R1

Calculate projected inference spend with prompt caching

Quick Workload Presets

Input Tokens / Req2,000
100100k200k
Output Tokens / Req500
5016k32k
Number of Requests1,000
125k50k
Cache Hit Rate0%
0% (No cache)50%90% (Max)
Total Input Spend

$1.1000

2.00M input tokens

Total Output Spend

$1.0950

0.50M output tokens

Estimated Total Cost

$2.20

Avg: $0.00220 / req

Rates: $0.55 in / $2.19 out per million tokens (DeepSeek Official API)Last verified: Aug 10, 2026

Deployment & Integration Options

Ollama (Local)local ollama
ollama run deepseek-r1:70b\n# or for smaller consumer laptops:\nollama run deepseek-r1:32b\nollama run deepseek-r1:14b
Hardware: 16GB - 64GB RAM/VRAMProvider Console
vLLM (Server / Enterprise)local vllm
vllm serve deepseek-ai/DeepSeek-R1 --trust-remote-code --tensor-parallel-size 8 --max-model-len 64000
Hardware: 8x 80GB GPUs (FP8)Provider Console
DeepSeek Official APIcloud api
curl https://api.deepseek.com/chat/completions \\\n  -H "Content-Type: application/json" \\\n  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \\\n  -d '{"model": "deepseek-reasoner", "messages": [{"role": "user", "content": "9.11 or 9.8, which is bigger?"}]}'
Hardware: Serverless APIProvider Console

Compare with Similar Models

Mistral AI

Codestral 2501

Mistral AI's dedicated code generation model with 256k context window and fill-in-the-middle (FIM) capabilities.

Compare DeepSeek-R1 vs Codestral 2501
DeepSeek

DeepSeek-V3

Frontier open-weights 671B MoE base and chat model with multi-head latent attention (MLA) and dual-pipe training.

Compare DeepSeek-R1 vs DeepSeek-V3
Meta AI

Llama 3.3 70B Instruct

Meta's open-weights 70B flagship matching Llama 3.1 405B capabilities on industry benchmarks at 1/5th the compute.

Compare DeepSeek-R1 vs Llama 3.3 70B Instruct

Data Accuracy & Verification Notice

AI model specifications, pricing records, and benchmark metrics published on this platform are compiled directly from authoritative sources (official provider documentation, research papers, and verified evaluation harnesses). Benchmark results reflect specific test harnesses and prompting methodologies; scores are not directly comparable across differing evaluation setups.