DeepSeek-R1 Release – Open-Source Model Matching OpenAI o1 Performance

TL;DR

DeepSeek released DeepSeek‑R1, an open‑source model whose performance on math, code, and reasoning benchmarks is on par with OpenAI o1, and made both the model and a suite of distilled variants freely available under the MIT license.

DeepSeek‑R1 Performance Claims

DeepSeek‑R1 is advertised as delivering "performance on par with OpenAI‑o1" across mathematics, programming, and logical reasoning tasks. The announcement includes benchmark graphics that compare DeepSeek‑R1 to OpenAI‑o1, indicating comparable scores on standard evaluation suites.

Open‑Source Availability and Licensing

  • The model weights, training code, and a technical report are released publicly.
  • DeepSeek‑R1 is distributed under the MIT License, allowing unrestricted commercial use, redistribution, and modification.
  • API outputs are explicitly permitted for downstream fine‑tuning and model distillation.

Distilled Model Family

DeepSeek also released six smaller distilled models derived from DeepSeek‑R1:

  • Two flagship sizes, 32 B and 70 B, are claimed to match the performance of OpenAI‑o1‑mini.
  • All six distilled variants are fully open‑source, providing the community with lighter‑weight alternatives for resource‑constrained environments.

Technical Highlights

  • Large‑scale reinforcement learning (RL) in post‑training – DeepSeek‑R1 applies RL after the primary pre‑training phase to improve reasoning capabilities.
  • Data efficiency – The model achieves a significant performance boost while requiring minimal labeled data for fine‑tuning.
  • Benchmark focus – Evaluation emphasizes mathematics, code generation, and multi‑step reasoning, where DeepSeek‑R1 reaches parity with OpenAI‑o1.
  • The full technical report is available at the GitHub repository: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf.

API Access and Pricing

  • The model can be invoked via the DeepSeek API by setting model=deepseek-reasoner.
  • Pricing structure (as of the announcement):
    • $0.14 per million input tokens when a cache hit occurs.
    • $0.55 per million input tokens on a cache miss.
    • $2.19 per million output tokens.
  • Detailed usage instructions are provided in the API guide: https://api-docs.deepseek.com/guides/thinking_mode.

Community Impact

  • By releasing both the flagship model and distilled variants under a permissive license, DeepSeek aims to empower the open‑source AI community and lower barriers to building advanced reasoning systems.
  • The open licensing and commercial‑friendly terms are positioned as a direct challenge to proprietary offerings, encouraging broader experimentation and adoption.

Visual Summary

The announcement includes several illustrative images:

  • An animated demo of DeepSeek‑R1 in action.
  • Benchmark charts comparing DeepSeek‑R1 to OpenAI‑o1.
  • A table summarizing the performance of the distilled models.
  • Pricing comparison graphics that contextualize DeepSeek‑R1 costs against competing services.

All statements are based solely on DeepSeek's official release dated 2025‑01‑20.

Sources

Related

  • Dispatch
    Open-R1: A Fully Open Reproduction of DeepSeek-R1

    Hugging Face has launched the Open-R1 project to systematically reconstruct the data and training pipeline of DeepSeek-R1, aiming to provide the open-source community with the missing datasets and code for reasoning models.

  • Dispatch
    Open-R1: Reproducing DeepSeek-R1 Reasoning Pipelines

    Hugging Face's Open-R1 project provides an open-source framework and datasets to reproduce the DeepSeek-R1 reasoning pipeline through distillation and reinforcement learning.

  • Dispatch
    OpenAI o1-preview release notes / what's new

    OpenAI has released o1-preview and o1-mini, a new series of reasoning models designed to solve complex problems in science, coding, and mathematics through extended thinking time.

  • Dispatch
    Open-R1 Update #1: Replicating DeepSeek-R1 Training and Synthetic Data

    Hugging Face provides a first progress update on the Open-R1 project, detailing the reproduction of DeepSeek-R1 evaluation scores, the integration of GRPO into TRL, and strategies for scaling synthetic reasoning data generation.

  • Dispatch
    OpenAI o1-mini release notes

    OpenAI has released o1-mini, a cost-efficient reasoning model optimized for STEM tasks that offers significantly lower latency and cost than o1-preview while maintaining high performance in math and coding.