DeepSeek-R1: Open-Weight Reasoning, Benchmark Results and a First Test

DeepSeek released R1, an open-weight reasoning model, in January 2025. I am excited because its reported results approach OpenAI o1 on several tests and its hosted API was priced lower in the comparison I show. Benchmark parity on selected tasks is not proof that the systems are equivalent in every capability.

DeepSeek’s chart compares R1, o1, a distilled R1 model, o1-mini and DeepSeek-V3 on mathematics, coding and general-knowledge evaluations. The reported results vary by benchmark: R1 leads on some and trails on others. These are developer-reported measurements under specific prompts and settings, not an independent universal ranking ([technical report](https://arxiv.org/abs/2501.12948)).

I see the availability of weights and a technical report as valuable for competition and research. I predict that open-weight reasoning models will rapidly follow closed models. The release supplies model weights and describes the approach, but “fully open source” should not be taken to imply that all pretraining data and every detail needed to reproduce the entire training run were released.

DeepSeek also released six smaller models distilled from R1 using Qwen and Llama bases, with sizes from 1.5B to 70B. The reported scores are promising, but the small models differ from the full R1 and from one another. DeepSeek says its repository and R1 weights use the MIT license; its [repository](https://github.com/deepseek-ai/DeepSeek-R1) also identifies the underlying Qwen and Llama licenses for the distilled variants, so check the relevant terms before use.

For the January 2025 launch, DeepSeek listed $0.14 per million input tokens on a cache hit, $0.55 on a cache miss and $2.19 per million output tokens. I compare them with the then-displayed o1 prices. These are historical provider rates, not current quotes; cache conditions, workload and the chosen model affect the comparison ([DeepSeek release](https://api-docs.deepseek.com/news/news250120/)).

My first prompt asks how many times the letter r appears in “strawberry.” The displayed reasoning goes back and forth and even briefly suggests an incorrect count, but the final answer is three. The readable trace is model-generated text; it is useful to inspect but should not be treated as a reliable account of an inner human-like thought process.

Next I ask about a marble in a glass that is inverted on a table and then moved to a microwave. The model considers whether the marble fits through the rim and concludes it remains on the table. I call this a good answer under the usual open-glass assumptions, although the exact outcome depends on the unstated handling and geometry.

At the time I note that the chat interface does not combine search and DeepThink in my test. I then request ten sentences ending in “apple,” and the displayed response satisfies the format. These few examples are informal observations, not a systematic comparison or proof that the model will be reliable on similar prompts.

DeepSeek’s paper describes R1-Zero as a model trained with large-scale reinforcement learning without an initial supervised fine-tuning stage. It starts from a pretrained base model, not from zero knowledge, and the authors report problems including readability and language mixing. R1 adds cold-start data and a multistage training pipeline to address these weaknesses.

I discuss group relative policy optimization, which compares a group of sampled outputs to estimate relative advantages instead of relying on a separately trained value critic in the described setup. This does not mean the training uses no rewards, evaluation or other supervision. DeepSeek also uses R1-generated examples to train smaller Qwen- and Llama-based models.

The paper shows an R1-Zero prompt format separating reasoning and answer sections, and describes a case where the model revisits its initial approach with additional reasoning. The authors call this an “aha moment.” One example illustrates a behavior observed under their training setup; it does not prove every visible revision is correct.

I compare reinforcement learning with AlphaGo’s advances in Go. The distinction matters: AlphaGo Zero learned from self-play without human game examples, while the famous move 37 occurred in the earlier AlphaGo match against Lee Sedol, not AlphaGo Zero ([DeepMind](https://deepmind.google/blog/alphago-zero-starting-from-scratch/)). I plan to explore the paper further and test the released models myself.