DeepSeek-R1
DeepSeek-R1 is a large language reasoning model developed by DeepSeek, focused on complex logical reasoning, mathematical computation, and code generation tasks. The model is trained using reinforcement learning and chain-of-thought techniques, exhibiting strong multi-step reasoning capabilities. It supports open-source deployment and can run on a variety of hardware platforms, representing a flagship version of DeepSeek's reasoning model series.
Core Facts
Reasoning models, represented by OpenAI o1, DeepSeek-R1, and QwQ, have chain-of-thought expansion as their core mechanism, generating a large number of intermediate reasoning steps before producing a final answer.
DeepSeek-V3 adopts a Mixture of Experts (MoE) architecture with a total parameter count of 671 billion, but only approximately 37 billion parameters are activated per inference.
According to official disclosure, DeepSeek's training cost was only approximately $6 million.
Timeline (last 90 days)
Old laptops without a discrete GPU, or mining cards such as the P106, are suitable for running the 1.5B or 7B versions of DeepSeek-R1; a P40 with 24G of VRAM can run the 32B version directly.
The UP host has seen cases in government departments and hospitals where tool calling was supported after fine-tuning DeepSeek-R1.
Reasoning tokens refer to the tokens consumed by the internal reasoning content generated by models with chain-of-thought capabilities (such as the OpenAI o series, DeepSeek-R1, etc.) before giving a final answer; they are usually not shown directly to the user but are counted toward billing and context window usage.
In the transfer of LLM reasoning capabilities, knowledge distillation usually refers to training a weak model using chain-of-thought answers generated by a strong model as supervision data, and DeepSeek-R1's distilled series falls into this category.
GPT's o1, DeepSeek-R1, and other reasoning models correspond to System 2 and first generate a chain of thought before answering, at the cost of high latency and large token consumption; decision models correspond to System 1 and map inputs directly to labels.
DeepSeek-R1, Claude Sonnet 5.5, and Claude Opus 5.5 scored 0.9375 on resolve but dropped to 0.875 on resolve-plus; DeepSeek-R1 leads the open-source group with these two scores.
DeepSeek-R1 scored zero across all 120 trials because it outputs raw <think> tokens into the response stream, exceeding the 256-token output budget before producing valid JSON.
When placing reasoning models behind strict micro-contracts, the thinking channel must be separated or stripped; otherwise the model will fail on protocol syntax.
DeepSeek-R1 repeatedly produces "wait" during reasoning, manifesting as linear enumeration of possibilities accompanied by repeated self-doubt.
Open-source models such as QwQ and DeepSeek-R1 are already able to compete on par with top closed-source commercial models on this navigation task.
32 more timeline events
All Facts (20)
Reasoning models, represented by OpenAI o1, DeepSeek-R1, and QwQ, have chain-of-thought expansion as their core mechanism, generating a large number of intermediate reasoning steps before producing a final answer.
90%VerifiedDeepSeek-V3 adopts a Mixture of Experts (MoE) architecture with a total parameter count of 671 billion, but only approximately 37 billion parameters are activated per inference.
90%VerifiedAccording to official disclosure, DeepSeek's training cost was only approximately $6 million.
80%VerifiedDeepSeek-R1 is a model focused on reasoning capabilities and possesses chain-of-thought ability through reinforcement learning training.
80%VerifiedThe OpenAI o1 series models were released in September 2024, representing a paradigm shift in LLMs from fast, intuitive responses to deep reasoning.
80%VerifiedDeepSeek R1 offers multiple distilled versions ranging from 1.5B to 671B parameters.
80%VerifiedOpenAI的o1系列模型和DeepSeek-R1采用了类似Chain-of-Thought的显式推理机制
80%VerifiedIn early 2025, DeepSeek-R1 achieved reasoning capabilities comparable to OpenAI o1 at a training cost of approximately $6 million.
80%Verified推理能力基于思维链（Chain-of-Thought, CoT）技术，模型在给出最终答案前生成内部逐步推导过程
80%VerifiedBefore outputting a final answer, a reasoning model's internal chain of thought may consume 3,000 to 20,000 tokens, far exceeding the approximately 500 tokens of an ordinary model.
75%VerifiedDeepSeek-R1 achieved success by running the GRPO algorithm at scale on two verifiable domains: code and mathematics.
75%VerifiedOpenAI's o-series reasoning models introduce a slow-thinking mechanism that performs internal chain-of-thought reasoning before outputting the final answer.
75%VerifiedMainstream open-source models such as Meta's Llama series, DeepSeek-R1, and Mistral mostly fall under the category of open-weight models.
70%VerifiedDeepMind research has shown that on certain mathematical reasoning tasks, smaller models can surpass models with larger parameter counts by increasing the amount of inference-time computation.
70%VerifiedGRPO is a reinforcement learning algorithm proposed by the DeepSeek team that eliminates the Critic Model from PPO, computing intra-group relative rewards as a baseline by generating a set of responses to the same question.
65%VerifiedReasoning models, represented by OpenAI's o series, DeepSeek-R1, and the Kimi K series, perform a chain-of-thought process before outputting, decomposing problems into multiple steps, performing self-verification, and deriving answers incrementally.
65%VerifiedThe weight files of models such as DeepSeek-R1 can be as large as hundreds of gigabytes.
65%VerifiedGoogle's Gemini 2.0 Flash Thinking and models such as DeepSeek-R1 have followed the test-time compute scaling technical path of o1.
65%VerifiedReasoning models (such as OpenAI's o1/o3 series, Anthropic's Claude with extended thinking, and DeepSeek-R1) output their thinking process (thinking tokens) before generating the final answer.
65%Unverified2024年底至2025年初，DeepSeek-V3和DeepSeek-R1相继发布
95%