DeepSeek open-sourced DeepSeek-R1, an LLM fine-tuned with reinforcement learning (RL) to improve reasoning capability. DeepSeek-R1 achieves results on par with OpenAI’s o1 model on several benchmarks, including MATH-500 and SWE-bench.
DeepSeek-R1 is based on DeepSeek-V3, a mixture of experts (MoE) model recently open-sourced by DeepSeek. This base model is fine-tuned using Group Relative Policy Optimization (GRPO), a reasoning-oriented variant of RL. The research team also performed knowledge distillation from DeepSeek-R1 to open-source Qwen and Llama models and released several versions of each; these models outperform larger models, including GPT-4, on math and coding benchmarks.
[DeepSeek-R1 is] the first step toward improving language model reasoning capabilities using pure reinforcement learning (RL). Our goal is to explore the potential of LLMs to develop reasoning capabilities without any supervised data, focusing on their self-evolution through a pure RL process…DeepSeek-R1…excels in a wide range of tasks, including creative writing, general question answering, editing, summarization, and more. Additionally, DeepSeek-R1 demonstrates outstanding performance on tasks requiring long-context understanding, substantially outperforming DeepSeek-V3 on long-context benchmarks.
To develop the model, DeepSeek started with DeepSeek-V3 as a base. They first tried fine-tuning it only with RL, and without any supervised fine-tuning (SFT), producing a model called DeepSeek-R1-Zero, which they have also released. This model exhibits strong reasoning performance, but ” powerful reasoning behaviors, it faces several issues. For instance, DeepSeek-R1-Zero struggles with challenges like poor readability and language mixing.”
To address this, the team used a short stage of SFT to prevent the “cold start” problem of RL. They collected several thousand examples of chain-of-thought reasoning to use in SFT of DeepSeek-V3 before running RL. After the RL process converged, they then collected more SFT data using rejection sampling, resulting in a dataset of 800k samples. This dataset was used for further fine-tuning and to produce the distilled models from Llama and Qwen.
DeepSeek evaluated their model on a variety of reasoning, math, and coding benchmarks and compared it to other models, including Claude-3.5-Sonnet, GPT-4o, and o1. DeepSeek-R1 outperformed all of them on several of the benchmarks, including AIME 2024 and MATH-500.
DeepSeek-R1 Performance. Image Source: DeepSeek-R1 Technical Report
Within a few days of its release, the LMArena announced that DeepSeek-R1 was ranked #3 overall in the arena and #1 in coding and math. It was also tied for #1 with o1 in “Hard Prompt with Style Control” category.
Django framework co-creator Simon Willison wrote about his experiments with one of the DeepSeek distilled Llama models on his blog:
Each response starts with a <think>…</think> pseudo-XML tag containing the chain of thought used to help generate the response. [Given the prompt] “a joke about a pelican and a walrus who run a tea room together”…It then thought for 20 paragraphs before outputting the joke!…[T]he joke is awful. But the process of getting there was such an interesting insight into how these new models work.
Andrew Ng’s newsletter The Batch wrote about DeepSeek-R1:
DeepSeek is rapidly emerging as a strong builder of open models. Not only are these models great performers, but their license permits use of their outputs for distillation, potentially pushing forward the state of the art for language models (and multimodal models) of all sizes.
The DeepSeek-R1 models are available on HuggingFace.